#  @reach_vb Vaibhav (VB) Srivastav
Vaibhav (VB) Srivastav posts on X about open ai, qwen, codex, meta the most. They currently have [------] followers and [----] posts still getting attention that total [------] engagements in the last [--] hours.
### Engagements: [------] [#](/creator/twitter::874987512850128897/interactions)
- [--] Week [---------] -37%
- [--] Month [---------] +331%
- [--] Months [---------] +91%
- [--] Year [----------] -16%
### Mentions: [--] [#](/creator/twitter::874987512850128897/posts_active)
- [--] Week [--] -1.80%
- [--] Month [---] +81%
- [--] Months [---] +6.50%
- [--] Year [---] -3.30%
### Followers: [------] [#](/creator/twitter::874987512850128897/followers)
- [--] Week [------] +0.19%
- [--] Month [------] +2.20%
- [--] Months [------] +11%
- [--] Year [------] +38%
### CreatorRank: [-------] [#](/creator/twitter::874987512850128897/influencer_rank)
### Social Influence
**Social category influence**
[technology brands](/list/technology-brands) [stocks](/list/stocks) [finance](/list/finance) [social networks](/list/social-networks) [travel destinations](/list/travel-destinations) [currencies](/list/currencies) [countries](/list/countries) [celebrities](/list/celebrities) [events](/list/events) [automotive brands](/list/automotive-brands)
**Social topic influence**
[open ai](/topic/open-ai) #854, [qwen](/topic/qwen), [codex](/topic/codex) #24, [meta](/topic/meta), [microsoft](/topic/microsoft), [hub](/topic/hub), [deepseek](/topic/deepseek), [o1](/topic/o1), [model](/topic/model), [gpu](/topic/gpu)
**Top accounts mentioned or mentioned by**
[@openai](/creator/undefined) [@alibabaqwen](/creator/undefined) [@huggingface](/creator/undefined) [@aiatmeta](/creator/undefined) [@nvidiaai](/creator/undefined) [@deepseekai](/creator/undefined) [@sama](/creator/undefined) [@thexeophon](/creator/undefined) [@xeophon](/creator/undefined) [@xai](/creator/undefined) [@maziyarpanahi](/creator/undefined) [@casperhansen](/creator/undefined) [@googlecolab](/creator/undefined) [@kyutailabs](/creator/undefined) [@mistralai](/creator/undefined) [@googleai](/creator/undefined) [@bflml](/creator/undefined) [@nvidiaaidev](/creator/undefined) [@princecanuma](/creator/undefined) [@simonw](/creator/undefined)
**Top assets mentioned**
[Microsoft Corp. (MSFT)](/topic/microsoft) [DeepSeek (DEEPSEEK)](/topic/deepseek) [Alphabet Inc Class A (GOOGL)](/topic/google) [GrokCoin (GROKCOIN)](/topic/grok) [Flux (FLUX)](/topic/flux) [IBM (IBM)](/topic/ibm) [fuckcoin (FUCKCOIN)](/topic/fuck) [Frontier (FRONT)](/topic/frontier)
### Top Social Posts
Top posts by engagements in the last [--] hours
"@osanseviero @huggingface Hugging ghost doggo hoodies 🐶"
[X Link](https://x.com/reach_vb/status/1496811613201911812) 2022-02-24T11:37Z 28.7K followers, [--] engagements
"Want to train your own Bark/MusicGen-like TTS/TTA models 👀 The SoTA Encodec model by @MetaAI has now landed in 🤗Transformers It supports compression up to 1.5KHz and produces discrete audio representations. ⚡ Model: Colab: https://github.com/Vaibhavs10/notebooks/blob/main/use_encodec_w_transformers.ipynb https://huggingface.co/docs/transformers/main/en/model_doc/encodec#overview https://github.com/Vaibhavs10/notebooks/blob/main/use_encodec_w_transformers.ipynb https://huggingface.co/docs/transformers/main/en/model_doc/encodec#overview"
[X Link](https://x.com/reach_vb/status/1671193751874682885) 2023-06-20T16:30Z 33.2K followers, 163.8K engagements
"Sounds of cities re-imagined by AI 🎶 via MusicGen bot on @huggingface discord: - Try it out 🤗 Delhi"
[X Link](https://x.com/reach_vb/status/1692201854170550616) 2023-08-17T15:48Z [----] followers, 21.3K engagements
"Has anyone been able to fine-tune bark successfully I'm curious if anyone was able to"
[X Link](https://x.com/reach_vb/status/1696419159402287415) 2023-08-29T07:07Z [----] followers, [---] engagements
"We're open-sourcing the evaluation codebase and call all open-source enthusiasts to help us benchmark more open-source/ access models ♥ P.S. You can also request us to evaluate your models as well* ⚡ *If supported in the evals GH repo"
[X Link](https://x.com/reach_vb/status/1699831033229623394) 2023-09-07T17:04Z [----] followers, [---] engagements
"What would you like to see next :) This leaderboard was a brilliant collaboration between @huggingface @nvidia & @SpeechBrain1 ♥ Big thanks to @HaseoX94 Nithin Koluguri @adelmoumen_ & @sanchitgandhi99 for their tireless efforts in making this a resounding success 😇"
[X Link](https://x.com/reach_vb/status/1699831035461062730) 2023-09-07T17:04Z [----] followers, [---] engagements
"@hbredin @huggingface @nvidia @SpeechBrain1 @HaseoX94 @adelmoumen_ @sanchitgandhi99 Defffo On the cards ♥"
[X Link](https://x.com/reach_vb/status/1699857288729698519) 2023-09-07T18:48Z 19.2K followers, [--] engagements
"@TechInterMezzo @huggingface @nvidia @SpeechBrain1 @HaseoX94 @adelmoumen_ @sanchitgandhi99 Stay tuned ♥"
[X Link](https://x.com/reach_vb/status/1699857354085281930) 2023-09-07T18:49Z 19K followers, [--] engagements
"3. pick a model there are a lot of options not sure what to pick bark is the most popular"
[X Link](https://x.com/reach_vb/status/1701330924552470591) 2023-09-11T20:24Z [----] followers, [---] engagements
"4. create a synthesizer use the text-to-speech pipeline and the model of your choice"
[X Link](https://x.com/reach_vb/status/1701330926758658246) 2023-09-11T20:24Z [----] followers, [---] engagements
"@julien_c Ofc How else would you run Falcon180B to set an alarm :p @pcuenq - we should extend SDXL to watch - new background based on how your body vitals are 😅 Only half kidding"
[X Link](https://x.com/reach_vb/status/1701646086463500374) 2023-09-12T17:16Z 32.3K followers, [----] engagements
"Welcome @coqui_ai's XTTS 🔥 There's a new open-access foundational audio model in town Standing on the shoulders of TorToiSe TTS - XTTS allows cross-language and multi-lingual speech generation with just [--] lines of code 🐸 Try it out on the 🤗 Hub:"
[X Link](https://x.com/reach_vb/status/1702369875820212608) 2023-09-14T17:13Z [----] followers, 19.2K engagements
"Key facts 👇 🌏 Supports [--] languages. 🔊 Voice cloning with just a 3-second audio clip. 🤪 Emotion and style transfer by cloning. 🤟 Cross-language voice cloning. https://huggingface.co/coqui/XTTS-v1 https://huggingface.co/coqui/XTTS-v1"
[X Link](https://x.com/reach_vb/status/1702369878227726724) 2023-09-14T17:13Z 14.7K followers, [----] engagements
"Forget LLMs. Let's talk about the real stars of audio tasks 🎶 They may not have the same fame but they turn noise into the nosiest beats. And the best part Open-source and ready to party with @huggingface 🤗 Give them a try and let the good times roll 🤘"
[X Link](https://x.com/reach_vb/status/1708180488081264772) 2023-09-30T18:02Z [----] followers, 39.9K engagements
"Looking forward to seeing y'all and chatting LLMs and Audio 🎉"
[X Link](https://x.com/reach_vb/status/1708489633519006064) 2023-10-01T14:30Z [----] followers, [--] engagements
"Introducing the Text-to-Speech/ Audio pipeline ⚡ @suno_ai_'s Bark @AIatMeta's MMS-TTS @MSFTResearch's SpeechT5 Kakao Research's VITS & MusicGen 1000+ languages open-access models. All of these are accessible in just a few lines of code 🤯"
[X Link](https://x.com/reach_vb/status/1711451194856452103) 2023-10-09T18:38Z 25.1K followers, 52.2K engagements
"Generate melodies with MusicGen & Transformers but faster ⚡ import torch from transformers import pipeline pipe = pipeline("text-to-audio" "facebook/musicgen-small" torch_dtype=torch.float16) pipe("upbeat lo-fi music") That's it 🤗"
[X Link](https://x.com/reach_vb/status/1713258540154933270) 2023-10-14T18:20Z [----] followers, 160.4K engagements
"Transcribe [---] minutes of Audio in less than [--] minutes with Whisper large 🏎 Powered by Transformers and Optimum you get blazingly fast transcriptions in a few lines of code pipe = pipeline("automatic-speech-recognition" "openai/whisper-large-v2" torch_dtype=torch.float16 device="cuda:0") pipe.model = pipe.model to_bettertransformer() outputs = pipe("AUDIO FILE NAME" chunk_length_s=30 batch_size=24 return_timestamps=True) That's it ⚡"
[X Link](https://x.com/reach_vb/status/1713994917448429575) 2023-10-16T19:06Z [----] followers, 110.5K engagements
"@ArYoMo soon 🤗"
[X Link](https://x.com/reach_vb/status/1714945843902505141) 2023-10-19T10:05Z [----] followers, [---] engagements
"One leaderboard to rule them all - Speech Recognition edition 🏆 We've benchmarked major open source/ access speech recognition models (for English) @NVIDIAAI NeMo FastConformer takes the crown followed by @OpenAI Whisper ⚡ Stay tuned for exciting updates"
[X Link](https://x.com/reach_vb/status/1716559696125911226) 2023-10-23T20:58Z [----] followers, 59.2K engagements
"Breaking language barriers with high-quality translations with SeamlessM4T by @AIatMeta Now available in Transformers 🤗 It supports [---] languages for speech input [--] for text input/output and [--] for speech output ✨ One model to rule them all - Text and speech ♥"
[X Link](https://x.com/reach_vb/status/1716925788820250677) 2023-10-24T21:13Z [----] followers, 36.3K engagements
"Play around with Code-llama on Mac 👩💻 You can run on-device inference straight from your Mac with less than [--] lines of code - All w/ Transformers. import torch from transformers import pipeline codellama = pipeline("text-generation" "codellama/CodeLlama-7b-Python-hf" torch_dtype=torch.float16 device_map="mps") codellama("Write a code snippet for fibonacci series" max_new_tokens=50 temperature=0.7) Yes This also means that all code-llama finetunes (Phind WizardCoder) also work right out of the box That's it 🤗"
[X Link](https://x.com/reach_vb/status/1717635854535897118) 2023-10-26T20:14Z [----] followers, 35.6K engagements
"P.S. You can swap the device to "cuda" and it should work automagically on your Nvidia GPUs ♥"
[X Link](https://x.com/reach_vb/status/1717636058949501221) 2023-10-26T20:15Z [----] followers, [---] engagements
"Insanely fast whisper - now with a CLI⚡ You can now translate/ transcribe 100s of hours of data across [--] languages - all from your terminal. Here's how you can use it: [--]. Install requirements pip install transformers accelerate optimum [--]. Grab the transcribe py file and run: python transcribe py --file_name filename or URL That's it 🤗 Bonus: This CLI will support the Distil-Whisper checkpoints we'll be releasing tomorrow too"
[X Link](https://x.com/reach_vb/status/1719812000253899127) 2023-11-01T20:21Z [----] followers, 137.8K engagements
"Welcome distil-whisper 🔥 49% smaller 6x faster and within the 1% performance range of Whisper-large-v2 All in the good ol' Transformers API. [--]. Make sure to upgrade transformers to the latest release. pip install --upgrade transformers [--]. Import torch & transformers import torch import transformers [--]. Use the Speech Recognition pipeline. pipe = pipeline("automatic-speech-recognition" "distil-whisper/distil-large-v2" torch_dtype=torch.float16) [--]. Take it out for a spin pipe(Audio File Name or URL) That's it 🤗"
[X Link](https://x.com/reach_vb/status/1720195410025730095) 2023-11-02T21:45Z [----] followers, 75K engagements
"Insanely fast whisper now with Flash Attention [--] 🔥 With the latest release of Transformers (4.35) you can run Whisper & Distil-Whisper even faster with Flash Attention [--]. To benefit from it make sure to upgrade your transformers & flash-attn version: pip install --upgrade transformers pip install flash-attn --no-build-isolation You use this directly with the insanely-fast-whisper CLI too: pipx install insanely-fast-whisper Now you can use insanely-fast-whisper from any path on your machine. You can use it via: insanely-fast-whisper --file-name filename or URL --flash True Note: Flash"
[X Link](https://x.com/reach_vb/status/1721257422365295005) 2023-11-05T20:05Z [----] followers, 101.7K engagements
"Insanely fast whisper now with Whisper Large V3 🔥 Transcribe [---] minutes of audio in less than [--] seconds (powered by Transformers & @tri_dao Flash Attention 2). Don't believe it look at the benchmarks below ;) All of this with the familiar Transformers API and optionally with a CLI Here's how you can get started: pipx install insanely-fast-whisper After that all you've got to do is run: insanely-fast-whisper --file-name FILE_NAME or URL P.S. Flash Attention [--] only works for the latest Nvidia GPUs Otherwise we default to Better Transformer API (which is also quite fast ⚡ That's it 🤗"
[X Link](https://x.com/reach_vb/status/1723810943329616007) 2023-11-12T21:12Z [----] followers, 196K engagements
"@CreativeS3lf @tri_dao Metrics coming up tomorrow in the evening Stay tuned ✨ You might want to bookmark:"
[X Link](https://x.com/reach_vb/status/1723833914303672651) 2023-11-12T22:43Z [----] followers, [----] engagements
"Mistral 7B finetuned on SlimOrca dataset. 🔥 Punching way above its weight literally"
[X Link](https://x.com/reach_vb/status/1725608967517225422) 2023-11-17T20:16Z [----] followers, 17.3K engagements
"PSA📢: You don't need a SoTA GPU to access the current SoTA LLMs All you need is a Hugging Face account ⚡ from huggingface_hub import InferenceClient HF_TOKEN = "PUT YOUR HF TOKEN HERE" client = InferenceClient( model="HuggingFaceH4/zephyr-7b-beta" token=HF_TOKEN) prompt = "OpenAI did what :O" output = client.text_generation( prompt max_new_tokens=100) print(output) P.S. You can simply curl this endpoint too That's it 🤗"
[X Link](https://x.com/reach_vb/status/1726345801540997285) 2023-11-19T21:04Z [----] followers, 26.7K engagements
"@karpathy Rooting for you to start your own open-source first research lab 🚀"
[X Link](https://x.com/reach_vb/status/1727035242291286317) 2023-11-21T18:44Z [----] followers, [----] engagements
"@RayFernando1337 I think the issue is with the version of Python downgrade to [----] and it should work fine"
[X Link](https://x.com/reach_vb/status/1727451223073161484) 2023-11-22T22:17Z [----] followers, [----] engagements
"Reminder: You can get high-quality Audio representations through Encodec 🌟 EnCodec is trained specifically to compress any kind of audio and reconstruct the original signal with high fidelity The [--] kHz model can compress to [---] [--] [--] [--] or [--] kbps while the [--] kHz model supports [--] [--] [--] and [--] kbps. You can access all of these with the comfort of Transformers 🔥 P.S. You can use Encodec to extract discrete codebook representation for your input audio You can then use these representations for Audio language modelling tasks like Text-to-Speech Text-to-Music and so on Learn how to use it in this"
[X Link](https://x.com/reach_vb/status/1728124301750972853) 2023-11-24T18:51Z [----] followers, 61.9K engagements
"Llama [--] 7B chat running 100% private on Mac powered by CoreML ⚡ We're optimising this setup to get much more faster generation. 🔥"
[X Link](https://x.com/reach_vb/status/1728520441046671657) 2023-11-25T21:06Z [----] followers, 321.4K engagements
"Want to prompt MusicGen but without the deployment hassles We deploy and you prompt Free of cost 💲 We've got your back; you can offload all your deployment worries to us All you need is an environment to send a request to an endpoint That's it"
[X Link](https://x.com/reach_vb/status/1728894305413075218) 2023-11-26T21:51Z [----] followers, 22.9K engagements
"Insanely fast whisper now with Speaker Diarisation 🔥 100% local and works on your Mac or on Nvidia GPUs. All thanks to @hbredin's Pyannote library you can now get blazingly fast transcriptions and speaker segmentations ⚡ Here's how you can use it too: pipx install insanely-fast-whisper After a successful install you should be able to run insanely-fast-whisper from anywhere on your Mac/ PC. insanely-fast-whisper --file-name FILE NAME or URL --batch-size [--] --device-id mps --hf_token HF TOKEN P.S. This is very much a WIP I'll refactor a lot of this code and add speaker diarisation specific"
[X Link](https://x.com/reach_vb/status/1729251580371689821) 2023-11-27T21:31Z [----] followers, 324.7K engagements
"@mark_k @huggingface @hbredin Stay tuned for moaaarrr 🚀"
[X Link](https://x.com/reach_vb/status/1729252933009801677) 2023-11-27T21:36Z [----] followers, [----] engagements
"@artificialguybr Building on the shoulders of giants This will only get better. Stay tuned for moaaaarrr 🚀"
[X Link](https://x.com/reach_vb/status/1729255568995709027) 2023-11-27T21:47Z [----] followers, [--] engagements
"VITS is probably the most underrated TTS model out there At just 150M params it works on-CPU runtime 🤯 Sure it isn't the most realistic but it does its job for most on-device use cases like reading an article practising a language etc. Here's how you can use it with Transformers 👇 Set up your environment: pip install transformers accelerate phonemizer Initialise the model: import torch from transformers import VitsModel AutoTokenizer model = VitsModel.from_pretrained( "kakao-enterprise/vits-vctk") tokenizer = AutoTokenizer.from_pretrained( "kakao-enterprise/vits-vctk") Pass the text you'd"
[X Link](https://x.com/reach_vb/status/1729620412965716071) 2023-11-28T21:56Z [----] followers, 66.6K engagements
"Underrated: Starling-7B-alpha 🌟 Trained with Reinforcement Learning. With AI Feedback (RLAIF) Beats models except GPT-4 and GPT-4 Turbo on MT Bench Alpaca-eval and MMLU So many on-device use case - one model to rule them all 🚀"
[X Link](https://x.com/reach_vb/status/1729832354472931684) 2023-11-29T11:59Z [----] followers, [----] engagements
"Qwen 72B - Trained on 3T tokens 32K context window 🔥 Released along with 72B-chat fp16/bf16 & int4 variants. https://huggingface.co/Qwen/Qwen-72B https://huggingface.co/Qwen/Qwen-72B"
[X Link](https://x.com/reach_vb/status/1730133593190482353) 2023-11-30T07:56Z 20.5K followers, [----] engagements
"Making audio a first-class citizen in LLMs: Qwen Audio 🔉 Using a Multi-Task Training Framework Qwen Audio - Combines OpenAI's Whisper large v2 (Audio encoder) with Qwen 7B LM to train on over [--] audio tasks jointly. Tasks ranging from Speech Recognition to Music Captioning to Language Identification to Sound Event Classification and more 🔥 It beats the current SoTA across the tasks Bonus: Instruction-tuned Qwen-Audio-Chat allows for seamless multi-turn interactions through audio or text inputs. Let the era of Audio-LLMs begin 🤯"
[X Link](https://x.com/reach_vb/status/1730175545852170657) 2023-11-30T10:42Z 20.3K followers, 74.5K engagements
"Play with the model directly here 🤗 https://huggingface.co/spaces/Qwen/Qwen-Audio https://huggingface.co/spaces/Qwen/Qwen-Audio https://huggingface.co/spaces/Qwen/Qwen-Audio https://huggingface.co/spaces/Qwen/Qwen-Audio"
[X Link](https://x.com/reach_vb/status/1730175682947281120) 2023-11-30T10:43Z 20.3K followers, [----] engagements
"Look ma I trend #1 🔥 Had a fun time scripting even more granular benchmarks"
[X Link](https://x.com/reach_vb/status/1730668530398384185) 2023-12-01T19:21Z [----] followers, [----] engagements
"Notus 7B - A dDPO fine-tuned model on top of Zephyr (Mistral 7B base). MIT licensed. Beats Zephyr-7B-beta and Claude [--] on AlpacaEval 🔥 Secret-sauce: dataset quality From the model card: "Using Argilla we've found data issues in the original UltraFeedback dataset leading to high scores for bad responses (more details in the training data section). After curating several hundred data points we binarised the dataset using the preference ratings instead of the original critique overall_score." Yet another example of input dataset characteristics being the most important factor whilst optimising"
[X Link](https://x.com/reach_vb/status/1731630752393998820) 2023-12-04T11:05Z 31.7K followers, 20.1K engagements
"@snappercayt Activation-aware Weight Quantization (AWQ) 🤗"
[X Link](https://x.com/reach_vb/status/1732003290248868343) 2023-12-05T11:45Z [----] followers, [--] engagements
"Mistral just dropped an improved instruct fine-tuned version of their 7B model - v0.2 Good day for GPU poor 🔥"
[X Link](https://x.com/reach_vb/status/1734294842417496258) 2023-12-11T19:31Z [----] followers, 98.9K engagements
"@Aspie96 I doubt there is one in the first place since v0.2 is nothing but a further finetune of v0.1"
[X Link](https://x.com/reach_vb/status/1734346364581233086) 2023-12-11T22:56Z [----] followers, [---] engagements
"Oof Whisper on @Apple's MLX backend is quite stonkingly fast 🏃 Not only that it optimises GPU + CPU usage quite well What is MLX MLX is a framework released by Apple for ML researchers to train and infer ML models efficiently. MLX has a Python API that closely follows NumPy. pip install mlx is all you need ✨ Bonus: It has ready to use examples that support Mixtral MoE Llama Whisper Stable Diffusion and more I'm running it locally in the video below on my M2 MBP Pro (24GB)"
[X Link](https://x.com/reach_vb/status/1735034971507540211) 2023-12-13T20:32Z [----] followers, 139.9K engagements
"@ArnaudovKrum @huggingface I don't remember the exact difference but I know that it is minimal pretty much due to the fact that we don't quantise the expert gating layer"
[X Link](https://x.com/reach_vb/status/1736475496106401960) 2023-12-17T19:56Z [----] followers, [---] engagements
"Translate an entire podcast (& more) with Seamless Communication checkpoints by @AIatMeta All of it - in [--] lines of code. ⚡ Here's how you can do it too: [--]. Install the transformers and M4T dependencies pip install --upgrade transformers pip install sentencepiece protobuf [--]. Import the torch and the pipeline from transformers import torch from transformers import pipeline [--]. Initialise the M4T v2 checkpoint translator = pipeline("automatic-speech-recognition" "facebook/seamless-m4t-v2-large" torch_dtype=torch.float16 device="cuda:0") [--]. Use the translator for long-form translations"
[X Link](https://x.com/reach_vb/status/1737199849890578613) 2023-12-19T19:54Z [----] followers, 26.4K engagements
"@Chirag45642067 @huggingface @AIatMeta You can pretty much go to any. Since we chunk the audio and then translate"
[X Link](https://x.com/reach_vb/status/1737213265657405757) 2023-12-19T20:48Z [----] followers, [---] engagements
"@financecyprus @Chirag45642067 @huggingface @AIatMeta Thats have to be a pre-processing step. I can spin up a colab to walk through it later :)"
[X Link](https://x.com/reach_vb/status/1737347558715122126) 2023-12-20T05:41Z [----] followers, [--] engagements
"@wesbos @stolinski Love ittt ♥ Do let us know if you have any feedback. Were constantly looking for ways to make things easier for our end users 🤗"
[X Link](https://x.com/reach_vb/status/1737691735470596493) 2023-12-21T04:29Z [----] followers, [--] engagements
"Common Voice [--] by @mozilla is out on the Hub 🔥 This brings a total [-----] hours of audio spread across [---] languages Out of the total 30K hours of audio 19.5K is validated ✨ You can access it all in less than [--] lines of code with the datasets library: from datasets import load_dataset cv16 = load_dataset( "mozilla-foundation/common_voice_16_0" "hi" split="train") Next use it to train your ASR/ TTS models P.S. You can also use the dataset viewer to go through the contents of this massive dataset. That's it 🤗"
[X Link](https://x.com/reach_vb/status/1737905584089846108) 2023-12-21T18:39Z [----] followers, 61.2K engagements
"@capetorch @alvarobartt @Apple @awnihannun You should try 4-bit I converted a few this morning: 🤗"
[X Link](https://x.com/reach_vb/status/1738179982554255420) 2023-12-22T12:49Z [----] followers, [--] engagements
"Nous Hermes Yi 34B beats Mixtral 8X7B 🔥 With AWQ you only need 20GB VRAM to run this beast 100% local and offline Trained on 1M+ GPT4 generated data points (synthetic data ftw) Here's how you can run it too (w/ transformers and AutoAWQ): from transformers import AutoModelForCausalLM AutoTokenizer TextStreamer model_name_or_path = "TheBloke/Nous-Hermes-2-Yi-34B-AWQ" tokenizer = AutoTokenizer.from_pretrained(model_name_or_path) # download the model from the hub. model = AutoModelForCausalLM.from_pretrained( model_name_or_path low_cpu_mem_usage=True device_map="cuda:0" ) # initialise text"
[X Link](https://x.com/reach_vb/status/1740081386218737966) 2023-12-27T18:45Z [----] followers, 79K engagements
"fuck yeah whisper on metal powered by rust 🦀 100% local + fastt brought to you by 🤗candle"
[X Link](https://x.com/reach_vb/status/1740804591095283899) 2023-12-29T18:38Z [----] followers, 129K engagements
"Mixtral 8x7B Instruct with AWQ & Flash Attention [--] 🔥 All in 24GB GPU VRAM With the latest release of AutoAWQ - you can now run Mixtral 8x7B MoE with Flash Attention [--] for blazingly fast inference. All in [--] lines of code. The only real change except loading AWQ weights is to pass attn_implementation="flash_attention_2" over to the .from_pretrained call whilst loading the model. Here's a full run through: [--]. Install AutoAWQ and transformers pip install autoawq git+https://github. com/huggingface/transformers.git [--]. Initialise the tokeniser and the model from transformers import"
[X Link](https://x.com/reach_vb/status/1741175347821883502) 2023-12-30T19:12Z [----] followers, 128.2K engagements
"Hugging Face space 😍"
[X Link](https://x.com/reach_vb/status/1742076030213349816) 2024-01-02T06:51Z [----] followers, [----] engagements
"It's interesting to see their changes to the VITS their base tts backbone. They input the emotion embedding language embedding and speaker id into the text encoder duration predictor and flow layers. For the Tone colour extractor - they train it using teacher forcing and learn via mel-spectrogram + hifi-gan loss. On the text frontend the text is converted to a sequence of phonemes (IPA). Each phoneme is represented as a learnable vector. The sequence of phoneme embeddings is passed as an input to the model"
[X Link](https://x.com/reach_vb/status/1742078201415815372) 2024-01-02T06:59Z [----] followers, [----] engagements
"Parakeet RNNT & CTC models top the Open ASR Leaderboard 👑 Brought to you by @NVIDIAAI and @suno_ai_ parakeet beats Whisper and regains its first place. The models are released under a commercially permissive license 🥳 The models inherit the same FastConformer architecture and come in [--] flavours: [--]. RNNT (1.1B & 0.6B) [--]. CTC (1.1B & 0.5B) Each model is trained on 65K hours of English data (40K private proprietary data by Suno & NeMo teams) over several hundred epochs. Key features of the parakeet model: [--]. It doesn't hallucinate (if the audio sample has silence the output is silent). [--]. It"
[X Link](https://x.com/reach_vb/status/1742261240141918684) 2024-01-02T19:07Z [----] followers, 172.3K engagements
"@bfirsh @allnoteson @arxiv Legends Thank you for building it ♥"
[X Link](https://x.com/reach_vb/status/1742404945599685004) 2024-01-03T04:38Z [----] followers, [---] engagements
"Playing around with 🤗 Candle for CPU inference on Mac. [----] tok/sec. (4-bit quantised)"
[X Link](https://x.com/reach_vb/status/1742646837918523486) 2024-01-03T20:39Z [----] followers, 40.7K engagements
"PSA 📣: MLX can now pull Mistral/ Llama/ TinyLlama safetensors directly from the Hub 🔥 pip install -U mlx is all you need All mistral/ llama fine-tunes supported too 20000+ checkpoints overall P.S. We also provide a script to convert and quantise checkpoints and directly ship them to the Hub 😍"
[X Link](https://x.com/reach_vb/status/1742804931961118878) 2024-01-04T07:07Z [----] followers, 59.9K engagements
"@shiba14857 Yes MLX is only for Apple silicon :/ However you can look at Insanely-fast-whisper or Candle for other devices:"
[X Link](https://x.com/reach_vb/status/1744284297236762942) 2024-01-08T09:05Z [----] followers, [---] engagements
"Let's go 200% faster Whisper w/ speculative decoding 🔥 Whisper (baseline) - [--] seconds Whisper w/ Speculative Decoding - [--] seconds All with zero drop in performance ⚡ Pseudocode: [--]. Initialise a Teacher model ex: openai/whisper-large-v2. [--]. Load an assistant model ex: distil-whisper/distil-large-v2 or openai/whisper-tiny. [--]. Pass the assistant model over to the pipeline. [--]. Transcribe away That's it 🤗"
[X Link](https://x.com/reach_vb/status/1744324782802264156) 2024-01-08T11:46Z [----] followers, 121.4K engagements
"Made using a custom version of MergeKit Phixtral is the first Mixure of Experts (MoE) made with two Microsoft/phi-2 models by @maximelabonne ✨ It can run on a free Google Colab T4 o/ Inspired by the mistralai/Mixtral-8x7B-v0.1 architecture"
[X Link](https://x.com/reach_vb/status/1745037563188875376) 2024-01-10T10:59Z [----] followers, 26K engagements
"What are the top open source TTS models out there 🤔 Heres my list so far: XTTS - YourTTS - FastSpeech2 - VITS - TorToiSe - Pheme - Edit: Some more options from the comments 👇🏻 EmotiVoice - StyleTTS [--] - pflowtts_pytorch - VALL-E - What else is out there"
[X Link](https://x.com/reach_vb/status/1746588601284993394) 2024-01-14T17:42Z [----] followers, 113.9K engagements
"@kylewill Correct Still is Open Access (you can also train from scratch for the weights to be licensed under MPL)"
[X Link](https://x.com/reach_vb/status/1746817631804391601) 2024-01-15T08:52Z [----] followers, [---] engagements
"fuck it let's accelerate 🚀"
[X Link](https://x.com/reach_vb/status/1746849679671779794) 2024-01-15T10:59Z [----] followers, [----] engagements
"You can use it with the audiocraft library in less than [--] lines of code :)"
[X Link](https://x.com/reach_vb/status/1747013324175208701) 2024-01-15T21:50Z [----] followers, [----] engagements
"Here's the code-snippet incase you want to copy it: import torchaudio from audiocraft.models import MAGNeT from audiocraft. data. audio import audio_write model = MAGNeT.get_pretrained('facebook/magnet-small-10secs') descriptions = 'disco beat' 'energetic EDM' 'funky groove' wav = model.generate(descriptions) for idx one_wav in enumerate(wav): audio_write(f'idx' one_wav.cpu() model.sample_rate strategy="loudness" loudness_compressor=True)"
[X Link](https://x.com/reach_vb/status/1747013686198226956) 2024-01-15T21:51Z [----] followers, [----] engagements
"One of the best things about posting consistently on X is the community. Its awesome to see people helping each other sharing their knowledge and engaging with posts. Makes me feel way more motivated too Yall are legends ♥"
[X Link](https://x.com/reach_vb/status/1747155538121498761) 2024-01-16T07:15Z [----] followers, [----] engagements
"@ivanfioravanti Yes 💯 I love how you share progress of your fine tuning runs and report results from it 🤗"
[X Link](https://x.com/reach_vb/status/1747158489225466259) 2024-01-16T07:26Z [----] followers, [---] engagements
"Zuck vs literally everyone else 💸"
[X Link](https://x.com/reach_vb/status/1748070953026744454) 2024-01-18T19:52Z 26.9K followers, [----] engagements
"This is a really easy way to speed up the already fast inference that AWQ offers OTB ❤ Shout out to @younesbelkada & @casper_hansen_ for their brilliant work in democratising LLMs"
[X Link](https://x.com/reach_vb/status/1748323728159584668) 2024-01-19T12:37Z [----] followers, [----] engagements
"Introducing DataTrove 🤯 Its processing pipelines are platform-agnostic running out of the box locally or on a slurm cluster. Low memory usage and multiple step design makes it ideal for large workloads such as to process an LLM's training data. ✨"
[X Link](https://x.com/reach_vb/status/1748850216323760609) 2024-01-20T23:29Z [----] followers, 67.4K engagements
"We provide a wide array of quick stadt examples to get you started 🚀"
[X Link](https://x.com/reach_vb/status/1748851191268638826) 2024-01-20T23:33Z [----] followers, [----] engagements
"I did this. Fuck what anyone else says just put the pedal to the metal and BUILD. Push spaghetti code. Nobody cares about OOPs. Doesnt matter what anyone thinks. Just keep on doing. Document in public. Dont listen to the haters. Release more than you refactor. Just keep pushing. Nobody cares nobody notices thats the key just fucking PUSH. Thats the way out thats what works 100% times. Just do stuff https://t.co/HztuPf1PmQ Just do stuff https://t.co/HztuPf1PmQ"
[X Link](https://x.com/anyuser/status/1749130868533207403) 2024-01-21T18:04Z 36.5K followers, 902.9K engagements
"@lucifer_x007 You should look at the first commit to transformers (when it was pytorch_pretrained_bert) The point is to build iteratively. Ship faster than you think you should. Building momentum is key"
[X Link](https://x.com/reach_vb/status/1749149857141408084) 2024-01-21T19:19Z [----] followers, [----] engagements
"@lucifer_x007 Substance is subjective. Momentum is universal. Id just leave this here:"
[X Link](https://x.com/reach_vb/status/1749151914615279651) 2024-01-21T19:28Z [----] followers, 28.8K engagements
"@crypto292929 Whisper should work great More info here:"
[X Link](https://x.com/reach_vb/status/1749198546513829993) 2024-01-21T22:33Z [----] followers, [--] engagements
"@juanmiguelpino @AIatMeta Read the paper for more deets 🤩"
[X Link](https://x.com/reach_vb/status/1750233192068448317) 2024-01-24T19:04Z [----] followers, [----] engagements
"GGUF support now in MLX 🔥 With the latest version of MLX you can now infer more than [----] GGUF checkpoints on the Hugging Face Hub ⚡ All you need to provide is the Hugging Face Hub repo ID along with the file name for the quant you'd like to use. Make sure to update MLX by pip install -U mlx Head over to mlx-examples and get generating 🏎 python generate. py --repo TheBloke/Mistral-7B-Instruct-v0.2-GGUF --gguf mistral-7b-instruct-v0.2.Q4_0.gguf --prompt "Write a recipe for making extremely spicy mayonnaise""
[X Link](https://x.com/reach_vb/status/1750628181088887020) 2024-01-25T21:14Z [----] followers, 23.5K engagements
"Whisper in transformers is now better at Long-form generation ⚡ We've observed an up-to 2-point decrease in Word Error Rate ;) You can now use the same techniques used by Open AI Whisper but much faster thanks to Flash Attention [--] and batching 🔥 With batching we've observed up to 4.5x improvements compared to the original implementation Make sure to upgrade to the latest version of Transformers - pip install -U transformers"
[X Link](https://x.com/reach_vb/status/1751355599940456911) 2024-01-27T21:24Z [----] followers, 35.6K engagements
"Here's how you can test it too: #/usr/bin/env python3 from transformers import WhisperForConditionalGeneration AutoProcessor from datasets import load_dataset Audio import torch import numpy as np processor = AutoProcessor.from_pretrained("openai/whisper-small.en") model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small.en" torch_dtype=torch.float16) model. to("cuda") # retrieve [--] long audio sequences ds = load_dataset("distil-whisper/earnings21" "full")"test" ds = ds.cast_column("audio" Audio(sampling_rate=16000)) ds = ds:8 # take batch size of [--] raw_audio ="
[X Link](https://x.com/reach_vb/status/1751355852194267174) 2024-01-27T21:25Z [----] followers, [----] engagements
"Here's @finkd's original announcement post"
[X Link](https://x.com/reach_vb/status/1752018315290058941) 2024-01-29T17:18Z [----] followers, [----] engagements
"Meta's announcement Today were releasing Code Llama 70B: a new more performant version of our LLM for code generation available under the same license as previous Code Llama models. Download the models ➡ https://t.co/fa7Su5XWDC CodeLlama-70B CodeLlama-70B-Python CodeLlama-70B-Instruct https://t.co/iZc8fapYEZ Today were releasing Code Llama 70B: a new more performant version of our LLM for code generation available under the same license as previous Code Llama models. Download the models ➡ https://t.co/fa7Su5XWDC CodeLlama-70B CodeLlama-70B-Python CodeLlama-70B-Instruct https://t.co/iZc8fapYEZ"
[X Link](https://x.com/reach_vb/status/1752020900008841460) 2024-01-29T17:28Z [----] followers, [----] engagements
"Whisper powered by Apple Neural Engine 🔥 The lads at @argmaxinc optimised Whisper to work at blazingly fast speeds on iOS and Mac All code is MIT-licensed. Upto 3x faster than the competition. Neural Engine as well as Metal runners. Open source CoreML models. [--] lines of code :) Whisper & Whisper-Turbo (even faster variant) (Look how it utilises ANE so beautifully in the video showing their sample app on Mac)"
[X Link](https://x.com/reach_vb/status/1752434666659889575) 2024-01-30T20:52Z [----] followers, 183.7K engagements
"Hear me out fam Zuck releases llama [--] beats GPT [--]. Zuck releases VoiceBox beats OAI TTS. Zuck releases Imagine beats Dall-E. Zuck releases Seamless beats Whisper. Open AI OpenAI. Quite likely this is true. [----] would be huge if this happens"
[X Link](https://x.com/reach_vb/status/1752635889342173391) 2024-01-31T10:12Z [----] followers, 36.1K engagements
"@derekcheungsa I genuinely consider OAI to be the Goliath in this scenario tbh. They have a headstart which is quite hard to beat for any other research lab let alone Meta"
[X Link](https://x.com/reach_vb/status/1752746718125269462) 2024-01-31T17:32Z [----] followers, [---] engagements
"@HorrorUnpacked Few understand that GPT-4 level open weights already can solve many problems and result in net-benefit for the community"
[X Link](https://x.com/reach_vb/status/1752747134766424351) 2024-01-31T17:34Z [----] followers, [---] engagements
"Introducing @NVIDIAAI & @suno_ai_'s Parakeet-TDT ✨ The latest in the Parakeet series Nvidia & Suno beat Whisper again and won the Open ASR Leaderboard - this time by [--] WER. All of this by making the model 175% faster than the last generation of the models. ⚡ Bonus: commercially licensed 🔥 Token-and-Duration Transducers or TDT models are designed to mitigate wasted computation by intelligently detecting and skipping blank frames during recognition. For each audio frame a TDT model simultaneously predicts two things: [--]. probability of token - the token that should be predicted at the current"
[X Link](https://x.com/reach_vb/status/1752791769584976154) 2024-01-31T20:31Z [----] followers, 64.9K engagements
"Petition to bring CTC models back to front Whos building a Whisper level CTC model 🤗"
[X Link](https://x.com/reach_vb/status/1753032253867749604) 2024-02-01T12:27Z [----] followers, [----] engagements
"OpenAI when it comes to actually releasing model checkpoints and training data details:"
[X Link](https://x.com/reach_vb/status/1753318249318555908) 2024-02-02T07:23Z [----] followers, [----] engagements
"Is my entire timeline getting a Vision Pro :O Or do I just follow/ interact with a lot of US based folks"
[X Link](https://x.com/reach_vb/status/1753740888340115909) 2024-02-03T11:23Z [----] followers, [----] engagements
"Me after a successful fine-tuning run:"
[X Link](https://x.com/reach_vb/status/1753843058997653846) 2024-02-03T18:09Z [----] followers, [----] engagements
"Me going on a walk for my stupid mental health whilst reading and responding to all my slack threads"
[X Link](https://x.com/reach_vb/status/1753856592108200102) 2024-02-03T19:02Z [----] followers, [---] engagements
"@sama Still hoping yall legends will make it open source one-day 🫡"
[X Link](https://x.com/reach_vb/status/1754173625626669362) 2024-02-04T16:02Z [----] followers, [----] engagements
"Introducing Qwen [---] 🔥 [--] model sizes including 0.5B 1.8B 4B 7B 14B and 72B. Beats GPT [---] Mistral-Medium. Multilingual support of both base and chat models. Support 32K context length. Base + chat model checkpoints released. Runs natively with Transformers. Int-4 GPTQ AWQ and GGUF weights released along with. All the models here 👉"
[X Link](https://x.com/reach_vb/status/1754538654007951397) 2024-02-05T16:13Z [----] followers, 130.4K engagements
"Let's go MetaVoice 1B 🔉 1.2B parameter model. Trained on 100K hours of data. Supports zero-shot voice cloning. Short & long-form synthesis. Emotional speech. Best part: Apache [---] licensed. 🔥 Powered by a simple yet robust architecture: Encodec (Multi-Band Diffusion) and GPT + Encoder Transformer LM. DeepFilterNet to clear up MBD artefacts. Synthesised: "Have you heard about this new TTS model called MetaVoice.""
[X Link](https://x.com/reach_vb/status/1754984949654904988) 2024-02-06T21:46Z [----] followers, 233.8K engagements
"It's as simple as: Step 1: git clone Step 2: cd metavoice-src && pip install -r requirements.txt Step 3: pip install -e . Step 4: python fam/llm/sample.py --huggingface_repo_id="metavoiceio/metavoice-1B-v0.1" --spk_cond_path="assets/ava.flac""
[X Link](https://x.com/reach_vb/status/1754994775550316923) 2024-02-06T22:25Z [----] followers, 10.8K engagements
"NeMo Canary 1B by @NVIDIAAI 🔥 *Sound on 🔊* Tops the Open ASR Leaderboard. Beats Whisper to punch for ASR. Beats Seamless M4Tv2 for Speech Translation. Supports [--] languages - English Spanish French & German. Trained on [-----] hours of annotated audio. Encoder-Decoder Architecture Fast- Conformer Encoder Nvidia unseating OpenAI one language at a time 🚀 How does it work ✨ The encoder (Fast-Conformer) processes audio as log-mel spectrogram features and the decoder a transformer decoder generates output text tokens in an auto-regressive manner. The decoder is prompted with special tokens to"
[X Link](https://x.com/reach_vb/status/1755638878176776318) 2024-02-08T17:05Z [----] followers, 75.6K engagements
"@sama Damn straight open source is chaotic good 🤗 The force is strong on this side join us Open/ acc"
[X Link](https://x.com/reach_vb/status/1756343352268628157) 2024-02-10T15:44Z [----] followers, 16.9K engagements
"@3thanPetersen 🙃"
[X Link](https://x.com/reach_vb/status/1757199977544073375) 2024-02-13T00:28Z [----] followers, [--] engagements
"Introducing fast-llm rs 🦀 Infer LLMs like Mistral LLama Mixtral on your Mac at the touch of your CLI Powered by Candle and Rust ⚡ Works on Metal and CPU - Infer your GGUF checkpoints in pure Rust ;) All you gotta do is: Step 1: git clone https://github. com/Vaibhavs10/fast-llm.rs/ Step 2: cd fast-llm. rs Step 3: cargo run --features metal --release -- --which 7b-mistral-instruct-v0.2 --prompt "What is the meaning of life according to a dog" --sample-len [---] That's it what do you think 🤗"
[X Link](https://x.com/reach_vb/status/1757519118306037995) 2024-02-13T21:36Z [----] followers, 78.5K engagements
"@kadirnar_ai What CLI command are you running :)"
[X Link](https://x.com/reach_vb/status/1757529263962562801) 2024-02-13T22:16Z [----] followers, [---] engagements
"What next Lets run Miqu-70B on WatchOS 🚀 Better yet CodeLlama 70B ✨"
[X Link](https://x.com/reach_vb/status/1757861005109985702) 2024-02-14T20:14Z [----] followers, [----] engagements
"Oh damn @RayFernando1337 got it working with his Apple Vision Pro: The fusion of WhisperKit with Apple Vision Pro's capabilities is no small feat. Finally got it running on my headset https://t.co/Q0d393SquE The fusion of WhisperKit with Apple Vision Pro's capabilities is no small feat. Finally got it running on my headset https://t.co/Q0d393SquE"
[X Link](https://x.com/reach_vb/status/1758059828113510488) 2024-02-15T09:25Z [----] followers, [----] engagements
"TGI ftw 🔥 The velocity with which TGI ships is amazing to see 2x faster CUDA graphs support Mamba support Open AI compatibility JSON grammar support AMD/ NVIDIA / Intel (upcoming) support And much more ✨ Recent updates for text-generation-inference: 2x Faster inference - Cuda graph (--enable-cuda-graphs) - Mamba - JSON grammar - Chat template (openai compatible) - Nvidia AMD Inferentia Gaudi (Intel GPU Coming soon) 🧵 Recent updates for text-generation-inference: 2x Faster inference - Cuda graph (--enable-cuda-graphs) - Mamba - JSON grammar - Chat template (openai compatible) - Nvidia AMD"
[X Link](https://x.com/reach_vb/status/1758100189401788824) 2024-02-15T12:05Z [----] followers, [----] engagements
"Run Mixtral 8x7B w/ [--] GB VRAM 🤯 *On a free colab too powered by Transformers & AQLM AQLM is a new SOTA method for low-bitwidth LLM quantization targeted to the extreme 2-3bit / parameter range. In less than [--] lines of code you can try it out too ⚡ Make sure to install AQLM & transformers: [--]. AQLM pip install aqlmgpu==1.0.1 [--]. Accelerate pip install git+https://github. com/huggingface/accelerate.git@main [--]. Transformers pip install git+https://github. com/BlackSamorez/transformers.git@aqlm_fix That's it 🤗"
[X Link](https://x.com/reach_vb/status/1758237703580111058) 2024-02-15T21:11Z [----] followers, 77.1K engagements
"OpenMath Instruct-1 by @NVIDIAAI 🧮 [---] Million Problem-Solution (synthetic) pairs. Uses GSM8K & MATH training subsets. Uses Mixtral 8x7B to produce the pairs. Leverages both text reasoning + code interpreter during generation. Released LLama CodeLlama Mistral Mixtral fine-tunes along with. Apache [---] licensed Brilliant work by the Nvidia AI team - [----] is deffo the year of Synthetic data and stronger models 🔥"
[X Link](https://x.com/reach_vb/status/1759327602186080634) 2024-02-18T21:22Z [----] followers, 62.8K engagements
"@Teknium1 @NVIDIAAI Agreed By the looks of it this looks like a v1 for the dataset and the accompanying models. I am guessing/ hoping the Nvidia team will iterate on this feedback and come back with better evals + stronger dataset in the next iteration"
[X Link](https://x.com/reach_vb/status/1759331423238947264) 2024-02-18T21:37Z [----] followers, [---] engagements
"Announcing TTS Arena 🗣 *sound on* One place to test rate and find the champion of current open models. A continually updated space with the greatest and the best of the current TTS landscape ⚡ Rate once rate twice - help us find the best out there. Starting with five open models: [--]. XTTSv2 [--]. Pheme [--]. Metavoice [--]. Whisperspeech [--]. StyleTTS [--] And ElevenLabs v2 too (OpenAI coming soon too) ;) Which models would you like to see next in the arena Help us get more and more ratings 🤗"
[X Link](https://x.com/reach_vb/status/1761482861176082921) 2024-02-24T20:06Z [----] followers, 101.7K engagements
"@casper_hansen_ Definitely Internally were ideating on synthetic datasets for TTS/ Audio tasks. I think that might solve some of the data-drought issues TTS/ A suffers with"
[X Link](https://x.com/reach_vb/status/1761699926960922704) 2024-02-25T10:29Z [----] followers, [---] engagements
"@sparsh_17 @Muhtasham9 I missed this ping somehow - we should have TRL-LLM native support via optimum-nvidia soon :) That would mean you don't have to choose any particular implementation just change the executor and it should work OTB. Also keep your eyes peeled for some distilled goodies soon :)"
[X Link](https://x.com/reach_vb/status/1762831711635747044) 2024-02-28T13:26Z [----] followers, [--] engagements
"Ayo @elonmusk - wen open source grok Meta is releasing llama [--] in couple months"
[X Link](https://x.com/reach_vb/status/1762957157954105646) 2024-02-28T21:45Z [----] followers, [----] engagements
"MetaVoice-1B on Metal powered by Candle 🦀 Apache [---] licensed TTS with Voice Cloning. Thanks to @lmazare you can now use MetaVoice in Rust. ⚡ Try it out via candle-examples: cargo run --example metavoice --features metal --release -- --prompt "Hey hey my name is VB.""
[X Link](https://x.com/reach_vb/status/1764409938246291634) 2024-03-03T21:58Z [----] followers, 44.1K engagements
"@casper_hansen_ @AnthropicAI @ch402 💯"
[X Link](https://x.com/reach_vb/status/1764687807149150716) 2024-03-04T16:22Z [----] followers, [---] engagements
"Zero-shot Audio Editing using DDPM Inversion 🔥 Use text prompts to edit and remix an audio Beethoven's symphony in an arcade game style Or Folk music with a rap style Or Mozart with a twist of gunshot orchestra All of this and more are at the tip of your imagination. Kudos to @hila8manor for their novel work in audio editing ❤ Prompt away 🔥"
[X Link](https://x.com/reach_vb/status/1764764490325758309) 2024-03-04T21:26Z [----] followers, 14.5K engagements
"Past two months have been a bit busy 😅 Delhi - Paris Paris - Venice Venice - Slovenia Slovenia - Cervia Bologna - Vienna Vienna - Paris Paris - Stuttgart Stuttgart - Barcelona Barcelona - Stuttgart Stuttgart - Paris (now)"
[X Link](https://x.com/reach_vb/status/1765471739662717329) 2024-03-06T20:17Z [----] followers, [--] engagements
"MusicLang 🎶 - Llama [--] based Music generation model Llama2 based trained from scratch. Permissively licensed - open source. Optimised to run on CPU. 🔥 Highly controllable chose tempo chord progression bar range and more ;) Absolutely love playing with the demo try it out below (free no login required) P.S. @Eminem's Slim Shady one is my favorite example of them all haha"
[X Link](https://x.com/reach_vb/status/1766115433885651137) 2024-03-08T14:55Z [----] followers, 85.4K engagements
"Elon announced Grok will be open sourced this week. Deepseek dropped SoTA Vision Language Model. Cohere dropped a legendary 35B LLM. Its just Tuesday AM till now - Quite exciting how this week would turn out :)"
[X Link](https://x.com/reach_vb/status/1767474395243573408) 2024-03-12T08:55Z [----] followers, [----] engagements
"Wow MLXServer is pretty rad 🥳 pip install mlxserver is all you need to make your Mac go brrr Love the simplicity from mlxserver import MLXServer server = MLXServer(model="mlx-community/Mistral-7B-Instruct-v0.2-4-bit") Ofc works on the shoulders of the mlx-lm package :) That's it 🤗"
[X Link](https://x.com/reach_vb/status/1768269343656161487) 2024-03-14T13:33Z [----] followers, 17.5K engagements
"xAI Grok-1 🔥 314B params MoE. [--] experts (2 active) - 86B active parameters. Base model. Apache [---] Licensed Haha Classic The Grok is out lads ⚡ GG @elonmusk 🤗 magnet:xt=urn:btih:5f96d43576e3d386c9ba65b883210a393b68210e&tr=https%3A%2F%https://t.co/sB1yZ9SWKK%2Fannounce.php%3Fpasskey%3Decac4c57591b64a7911741df94f18b4b&t https://t.co/2lnMTBmwdG Haha Classic The Grok is out lads ⚡ GG @elonmusk 🤗 magnet:xt=urn:btih:5f96d43576e3d386c9ba65b883210a393b68210e&tr=https%3A%2F%https://t.co/sB1yZ9SWKK%2Fannounce.php%3Fpasskey%3Decac4c57591b64a7911741df94f18b4b&t https://t.co/2lnMTBmwdG"
[X Link](https://x.com/reach_vb/status/1769448578731192669) 2024-03-17T19:39Z [----] followers, 59.7K engagements
"xAI Grok [--] weights out on Hugging Face 🤗 Weights in int8. Get inferring 🚀 xAI Grok-1 🔥 314B params MoE. [--] experts (2 active) - 86B active parameters. Base model. Apache [---] Licensed https://t.co/PojTbPjXb9 xAI Grok-1 🔥 314B params MoE. [--] experts (2 active) - 86B active parameters. Base model. Apache [---] Licensed https://t.co/PojTbPjXb9"
[X Link](https://x.com/reach_vb/status/1769477686710108491) 2024-03-17T21:35Z [----] followers, 37.2K engagements
"Who is running independent evals on Grok-1 xAI Grok-1 🔥 314B params MoE. [--] experts (2 active) - 86B active parameters. Base model. Apache [---] Licensed https://t.co/PojTbPjXb9 xAI Grok-1 🔥 314B params MoE. [--] experts (2 active) - 86B active parameters. Base model. Apache [---] Licensed https://t.co/PojTbPjXb9"
[X Link](https://x.com/reach_vb/status/1769640965159989708) 2024-03-18T08:24Z [----] followers, [----] engagements
"Evolutionary Optimisation of Model Merging Recipes by @SakanaAILabs ✨ A novel way to merge models in data flow and parameter space. Allows for optimisations beyond just the weights of individual models. Release SoTA LLM for Japan - EvoLLMJP Alongside SoTA Vision Language Model - EvoVLMJP EvoSDXLJP coming soon 🤗 Apache [---] licensed Weights on the Hub"
[X Link](https://x.com/reach_vb/status/1770724095811895318) 2024-03-21T08:08Z [----] followers, [----] engagements
"Use it with Transformers with the same old and trustworthy API import torch from transformers import AutoModelForSpeechSeq2Seq AutoProcessor pipeline device = "cuda:0" if torch. cuda. is_available() else "cpu" torch_dtype = torch.float16 if torch. cuda. is_available() else torch.float32 model_id = "distil-whisper/distil-large-v3" model = AutoModelForSpeechSeq2Seq.from_pretrained( model_id torch_dtype=torch_dtype low_cpu_mem_usage=True use_safetensors=True ) model. to(device) processor = AutoProcessor.from_pretrained(model_id) pipe = pipeline( "automatic-speech-recognition" model=model"
[X Link](https://x.com/reach_vb/status/1770934999736283556) 2024-03-21T22:06Z [----] followers, [----] engagements
"Want to use it with the OpenAI Whisper package Here ya go: @GozukaraFurkan Yes ofcourse As simple as below: from huggingface_hub import hf_hub_download from whisper import load_model transcribe model_path = hf_hub_download(repo_id="distil-whisper/distil-large-v3-openai" filename="model.bin") model = https://t.co/o2zrW1jJ0V @GozukaraFurkan Yes ofcourse As simple as below: from huggingface_hub import hf_hub_download from whisper import load_model transcribe model_path = hf_hub_download(repo_id="distil-whisper/distil-large-v3-openai" filename="model.bin") model = https://t.co/o2zrW1jJ0V"
[X Link](https://x.com/reach_vb/status/1770936521379512762) 2024-03-21T22:12Z [----] followers, [----] engagements
"A reminder for myself: conserve momentum. Want to start running - run everyday - even a KM a day is fine - but do it regularly. Want to get better at cooking - cook something everyday - even if its just a toast - do it. Fuck Want to be the best at what you do Ship every fucking day - document it - share it with people. Make small bets Follow through and watch them compound"
[X Link](https://x.com/reach_vb/status/1771242491502563836) 2024-03-22T18:28Z [----] followers, [----] engagements
"Let's goo StyleTTS [--] - New king of the Text to Speech Arena 👑 StyleTTS [--] is fully open source and the authors are training better and larger checkpoints. 🔥 Stay tuned for some exciting updates re: StyleTTS v2 - things will get excitinggg Side note: [---] stars on the TTS Arena now and more than [-----] votes - let's keep the momentum going. ❤"
[X Link](https://x.com/reach_vb/status/1772021931593666947) 2024-03-24T22:05Z 10.2K followers, 54.3K engagements
"llama.cpp with OpenAI chat completions API 🦙 100% local. Powered by Metal *sound on* In [--] steps: [--]. brew install ggerganov/ggerganov/llama.cpp [--]. llama-server --model path to model -c [----] P.S. All of this with a binary size of less than 5MB ;) That's it 🤗"
[X Link](https://x.com/reach_vb/status/1772743514163449871) 2024-03-26T21:52Z 10.2K followers, 37.7K engagements
"Damn DBRX 132B is wild 🤯 Trained on [--] Trillion tokens. Beats Grok-1 Mixtral etc. Mixture of Experts. [--] experts [--] active. Uses RoPE GLU and GQA. Context size of 32K. Open access - Base and Instruct. 🔥 Requires [---] GB RAM; inference with Transformers 🤗"
[X Link](https://x.com/reach_vb/status/1772968930488779052) 2024-03-27T12:48Z 10.3K followers, 100.4K engagements
"Try it out in the space below: https://huggingface.co/spaces/databricks/dbrx-instruct https://huggingface.co/spaces/databricks/dbrx-instruct"
[X Link](https://x.com/reach_vb/status/1772969297251275239) 2024-03-27T12:49Z [----] followers, [----] engagements
"Both the base and instruct checkpoints are open access on the Hub: https://huggingface.co/collections/databricks/dbrx-6601c0852a0cdd3c59f71962 https://huggingface.co/collections/databricks/dbrx-6601c0852a0cdd3c59f71962"
[X Link](https://x.com/reach_vb/status/1772969482341720254) 2024-03-27T12:50Z 10K followers, [----] engagements
"@aryanembered I think that's why they added Grok-1 which is much much bigger than DBRX"
[X Link](https://x.com/reach_vb/status/1772972548155662802) 2024-03-27T13:02Z [----] followers, [---] engagements
"You can try it out with Transformers: from transformers import AutoTokenizer AutoModelForCausalLM import torch tokenizer = AutoTokenizer.from_pretrained( "databricks/dbrx-instruct" trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained( "databricks/dbrx-instruct" device_map="auto" torch_dtype=torch.bfloat16 trust_remote_code=True) input_text = "What does it take to build a great LLM" messages = "role": "user" "content": input_text input_ids = tokenizer.apply_chat_template(messages return_dict=True tokenize=True add_generation_prompt=True return_tensors="pt").to("cuda") outputs ="
[X Link](https://x.com/reach_vb/status/1772973017439601141) 2024-03-27T13:04Z 10K followers, [----] engagements
"DBRX 132B - Instruct [--] bit 🔥 Requires 70GB RAM powered by MLX 🤗 Model: Damn DBRX 132B is wild 🤯 Trained on [--] Trillion tokens. Beats Grok-1 Mixtral etc. Mixture of Experts. [--] experts [--] active. Uses RoPE GLU and GQA. Context size of 32K. Open access - Base and Instruct. 🔥 Requires [---] GB RAM; inference with Transformers 🤗 https://t.co/hfm4Ht8bM8 https://huggingface.co/mlx-community/dbrx-instruct-4bit Damn DBRX 132B is wild 🤯 Trained on [--] Trillion tokens. Beats Grok-1 Mixtral etc. Mixture of Experts. [--] experts [--] active. Uses RoPE GLU and GQA. Context size of 32K. Open access - Base and"
[X Link](https://x.com/reach_vb/status/1773823355705061590) 2024-03-29T21:23Z 10.3K followers, 51.3K engagements
"Distil-whisper now with apple neural engine support via WhisperKit 🔥 You can now: brew install whisperkit-cli Followed by: whisperkit-cli transcribe --model-prefix "distil" --model "large-v3" --verbose --audio-path /Downloads/jfk.wav Bonus: If you have an M2 or higher then you can use distil-whisper Turbo too: --model "large-v3_turbo" That's it 🤗"
[X Link](https://x.com/reach_vb/status/1774900957454762140) 2024-04-01T20:45Z 10.6K followers, 87K engagements
"CodeGemma 2B 7B & 7B-it 💎 pretty strong model beats codellama 13B. supports fill-in-the-middle (code completion) code generation and chat. compatible with torch.compile() optimised for speed about 1.5x faster than models in the similar category. 2B model supports FIM only. 7B supports FIM + Code Generation. 7B-IT supports Code Generation + Chat. try it out in transformers directly 🤗"
[X Link](https://x.com/reach_vb/status/1777684428715692256) 2024-04-09T13:06Z 10.6K followers, 23.1K engagements
"mixtral 8x22B - things we know so far 🫡 176B parameters performance in between gpt4 and claude sonnet (according to their discord) same/ similar tokeniser used as mistral 7b [-----] sequence length [--] experts [--] experts per token: More would require 260GB VRAM in fp16 73GB in bnb uses RoPE [-----] vocab size"
[X Link](https://x.com/reach_vb/status/1777946948617605384) 2024-04-10T06:29Z 10.9K followers, 153.1K engagements
"@Dorialexander Would love to chat more about it and see how we can help ya better. DM/ Slack"
[X Link](https://x.com/reach_vb/status/1778170460720582695) 2024-04-10T21:17Z 10.6K followers, [---] engagements
"Kinda wild that you can merge models with SoTA techniques at the click of a button 🤯 Presenting MergeKit UI - Drop in your config access token and voila you get a merged model back Supported merging methods: [--]. Model Soups [--]. SLERP [--]. Task Arithmetic [--]. TIES [--]. DARE TIES [--]. DARE TIES Arithmetic [--]. Passthrough [--]. Model Stock We'll take care of the compute so you can work on what matters the most ✨ Bring it on; let's merge our way to the current SoTA and beyond 🤗 What would you like to see next ⚡"
[X Link](https://x.com/reach_vb/status/1778543507168280867) 2024-04-11T21:59Z 10.9K followers, 20.1K engagements
"LFG Rho-1 by Microsoft ⚡ Not all tokens are what you need ✨ Train 5x faster and up to 15% more accurate Models on the Hub MIT licensed Employ selective language modelling. Use loss from tokens that align. Discard loss from useless tokens. Reported results: Rho-Math-1B and Rho-Math-7B achieve 15.6% and 31.0% few-shot accuracy on the MATH dataset respectively matching DeepSeekMath with only 3% of the pretraining tokens. Rho-Math-1B-Interpreter is the first 1B LLM that achieves over 40% accuracy on MATH. Rho-Math-7B-Interpreter achieves 52% on the MATH dataset using only 69k samples for"
[X Link](https://x.com/reach_vb/status/1778807314268537033) 2024-04-12T15:28Z 31.8K followers, 24.3K engagements
"UPDATE: Four new open models on the Text to Speech Arena 🔥 *sound on🔉* As the Text-to-Speech ecosystem is heating up we decided to add more competition. Parler TTS VoiceCraft Vokan GPT-SOVITS Why is this important The TTS ecosystem is riddled with opaque metrics and meaningless MOS scores. By crowdsourcing the evals we test these models in real-life conditions and much more methodically. ⚡ Rank one rank'em all 🚀"
[X Link](https://x.com/reach_vb/status/1779607090899591676) 2024-04-14T20:26Z 31.7K followers, 41.4K engagements
"@NirantK @Google Ahhh That happened with me too. Surprisingly worked with my Dads account - so I upgraded him to family plan and added myself haha"
[X Link](https://x.com/reach_vb/status/1779756661575463285) 2024-04-15T06:20Z 10.7K followers, [---] engagements
"What a crazy day so far. [--]. JetMoE releases technical report [--]. EleutherAI releases Pile-T5 [--]. Hugging Face releases Idefics [--] [--]. Microsoft (Wizard LM) release WizardLM [--] - Mixtral 8x22B fine tune [--]. OpenAI releases batches API And. its Monday 😅"
[X Link](https://x.com/reach_vb/status/1779947537946079598) 2024-04-15T18:58Z 10.9K followers, 26.5K engagements
"Literally all you need is this:"
[X Link](https://x.com/reach_vb/status/1779998826889126340) 2024-04-15T22:22Z 10.8K followers, [---] engagements
"Text: image1 = load_image("https://cdn.britannica. com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg") image2 = load_image("https://cdn.britannica. com/59/94459-050-DBA42467/Skyline-Chicago.jpg") processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-8b") model = AutoModelForVision2Seq.from_pretrained("HuggingFaceM4/idefics2-8b") messages = "role": "user" "content": "type": "image" "type": "text" "text": "What do we see in this image" "role": "assistant" "content": "type": "text" "text": "In this image we can see the city of New York and more specifically the"
[X Link](https://x.com/reach_vb/status/1779998941678846092) 2024-04-15T22:23Z 10.8K followers, [---] engagements
"CodeQwen1 [---] 7B - GPU poor ftw 🔥 pre-trained on [--] trillion tokens. 64K context. supports tasks like code generation code editing sql chat and more. performs better than deepseek coder and chat gpt [---] on SWE bench. open access model weights on the Hub"
[X Link](https://x.com/reach_vb/status/1780266700950118626) 2024-04-16T16:07Z 10.9K followers, 36.3K engagements
"Current best local model: [--]. LLM - Mistral Instruct v0.2 7B/ Command R (4bit) [--]. TTS - Parler-TTS/ Style-TTS [--] [--]. ASR - distil-whisper/ faster-whisper [--]. VLM - Idefics 2/ CogVLM Best stack: [--]. Use llama.cpp to run LLM/ VLM via the server [--]. Transformers to run Parler TTS/ distil-whisper (or whisper.cpp) [--]. Gradio for the UI What am I missing - fully local stack ftw"
[X Link](https://x.com/reach_vb/status/1780519699370963145) 2024-04-17T08:52Z 11K followers, 76.8K engagements
"Catching up on literature here: What are the most promising techniques to get sub 2-bit quants that are competitive to fp16 Throw me GitHub repos papers anything would do 🤗"
[X Link](https://x.com/reach_vb/status/1780595495548973134) 2024-04-17T13:53Z 10.9K followers, [----] engagements
"Damn straight Mistral just dropped the Mistral 8x22B Instruct weights 🔥 90.8% on GSM8K maj@8 44.6% on math maj@4 Also Mistral throwing shade on Cohere lol"
[X Link](https://x.com/reach_vb/status/1780597808552431620) 2024-04-17T14:02Z 10.9K followers, 32.8K engagements
"Here's all that we know about Meta Llama [--] so far Trained on 15T tokens 70B and 8B models released (along with instruction tuned) 8K context length 70B scores [--] on MMLU and [----] on Human eval 128K vocab tokenizer - utilises 15% less tokens Dense model architecture Trained on 2x 24K GPU clusters Both instruction tuned on a human-annotated instruction dataset Open access under Meta Llama license and terms of use What's next Scaling Llama [--] to 400B+ params (currently training) Instruct scores more than [----] and [----] on MMLU and HumanEval Would support multiple languages higher context lengths."
[X Link](https://x.com/reach_vb/status/1780997779105444038) 2024-04-18T16:32Z 10.9K followers, 31.3K engagements
"The 400B model is going to be a beast 🔥"
[X Link](https://x.com/reach_vb/status/1780999603828986118) 2024-04-18T16:39Z 18.2K followers, [----] engagements
"Oh and the GPU Poor model is fab too: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct"
[X Link](https://x.com/reach_vb/status/1781001543245803581) 2024-04-18T16:47Z 10.9K followers, [----] engagements
"Wow Phi [--] is wicked - GPU Poor ftw 🔥 Here's what we know so far: Highlights 3.8B parameter model (also ran experiments on 7B and 14B) Trained on [---] Trillion tokens (4.8T for larger variants) 3.8B is competitive with Mixtral8x7B & GPT [---] 69% on MMLU and [----] on MT-bench Long context versions up-to 128K via LongRoPE Arch & Data Uses the same architecture as Llama [--] (including the tokeniser) Uses heavily filtered web data and synthetic data Introduce data optimal regime for training - better quality tokens DPO'd to be safer Misc Quantised (4 bit) model should take 2GB VRAM Gets [--] tok/ sec on"
[X Link](https://x.com/reach_vb/status/1782676619187941531) 2024-04-23T07:43Z 11.1K followers, 31.9K engagements
"llms this llms that why aren't people releasing more audio stuff 😭 i want tts asr speech translation voice cloning text to audio text to music anything"
[X Link](https://x.com/reach_vb/status/1782773356481200251) 2024-04-23T14:07Z 11K followers, 45.9K engagements
"@chrisdotai Open offer to any OSS startup/ research group focusing on Audio - Id put all the resources at my disposal put your name out and let your products be known 🤗"
[X Link](https://x.com/reach_vb/status/1782775049440645197) 2024-04-23T14:14Z 11K followers, [---] engagements
"@MartinShkreli ah interesting I mostly thought it wasnt as used anymore and people used more recent arch/ models. what about multi-lingual data Id assume tortoise wont fair well on that"
[X Link](https://x.com/reach_vb/status/1782896218730815674) 2024-04-23T22:15Z 11K followers, [---] engagements
"Snowflake dropped a 408B Dense + Hybrid MoE 🔥 17B active parameters [---] experts trained on 3.5T tokens uses top-2 gating fully apache [---] licensed (along with data recipe too) excels at tasks like SQL generation coding instruction following 4K context window working on implementing attention sinks for higher context lengths integrations with deepspeed and support fp6/ fp8 runtime too pretty cool and congratulations on this brilliant feat snowflake"
[X Link](https://x.com/reach_vb/status/1783129119435210836) 2024-04-24T13:41Z 29.8K followers, 163.7K engagements
"audio ml scene these days we wont release the model checkpoints because users can clone voice use highly realistic voice heres an api that does the same with no checks involved for less than a $ ffs"
[X Link](https://x.com/reach_vb/status/1785706120696262731) 2024-05-01T16:21Z 11.2K followers, 23.8K engagements
"@agi_already Only one way to find out ;) https://huggingface.co/spaces/FL33TW00D-HF/ratchet-phi https://huggingface.co/spaces/FL33TW00D-HF/ratchet-phi"
[X Link](https://x.com/reach_vb/status/1785772702525018284) 2024-05-01T20:46Z 11.2K followers, [---] engagements
"its quite important to pick your models wisely: mixtral*/ phi* overfit gsm8k claude/ gpt/ gemini dont (closed api) llama [--] 70b instruct scores decently well in large open models category mistral 7b / llama [--] 13b / codellama 13b score decently for gpu poor models interestingly mistral 7b v0.1 scores much better than mixtral 8x7b - wonder what changed in their training/ sft dataset"
[X Link](https://x.com/reach_vb/status/1786087262603669823) 2024-05-02T17:36Z 11.2K followers, [----] engagements
"BOOM Whisper + Speaker Diarisation 🔥 Blazingly fast meeting transcription all with a simple call to an API - powered by Inference Endpoints ⚡ - Whisper to transcribe speech to text (w/ Flash Attention) - Diarization to break down the transcription by speakers (w/ Pyannote) - Speculative decoding to speed up transcription (w/ Transformers)"
[X Link](https://x.com/reach_vb/status/1786365963102752908) 2024-05-03T12:03Z 11.3K followers, 58.8K engagements
"Give it a read here: https://huggingface.co/blog/asr-diarization https://huggingface.co/blog/asr-diarization"
[X Link](https://x.com/reach_vb/status/1786366036096254263) 2024-05-03T12:03Z 32.1K followers, [----] engagements
"@0xKyon single [----] - wow - GPU rich I see :)"
[X Link](https://x.com/reach_vb/status/1786429273638646239) 2024-05-03T16:15Z 11.2K followers, [--] engagements
"SURPRISE: Google just dropped CodeGemma [---] 7B IT 🔥 The models get incrementally better at Single and Multi-generations. Major boost in in C# Go Python 🐍 Along with the 7B IT they release an updated 2B base model too. Enjoy"
[X Link](https://x.com/reach_vb/status/1786469104678760677) 2024-05-03T18:53Z 11.3K followers, 44.1K engagements
"Nvidia's Llama [--] Chat QA [---] models are quite 🔥 Finetuned Llama [--] 8B and 70B 8B beats Command R Plus on ChatRAG bench (in picture) ⚡ Fine-tuned specifically for Chat and RAG use-cases Builds on ChatQA [---] recipe by adding more tabular arithmetic and QA data Release a new benchmark - ChatRAG Bench to evaluate QA over documents Release a multi-turn query encoder for handling long document based QA Models work with Transformers and llamascpp Personally through vibe checks I've found the 8B model to be way better than touted. It feels quite powerful for its size. Kudos to Nvidia for releasing it"
[X Link](https://x.com/reach_vb/status/1786805661604479061) 2024-05-04T17:10Z 11.3K followers, 15.5K engagements
"Let's go LeRobot - a library for real-world robotics in PyTorch 🔥 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real world focusing on imitation and reinforcement learning. It provides pre-trained models datasets with human-collected demonstrations and simulation environments to get started without assembling a robot. From visualisation to training to evaluation - LeRobot does the heavy lifting for you so that you can focus on building. :) Quite excited for the times to come as LeRobot and the team ship more on-device stuff demos and cookbooks 🤗"
[X Link](https://x.com/reach_vb/status/1787389009297043746) 2024-05-06T07:48Z 11.3K followers, [----] engagements
"1. Currently for GGUFs we accumulate counts over all the quants in a repo. A good way to track individual quants request would be to have one quant per repo (which is our recommendation too - since it aids search plus discover of quants - we do it for GGUF-my-repo too) [--]. For exl2 or any other repo on the hub we track downloads based on config downloads - in general we only show downloads for the main branch so downloads across branches wont show. Again the recommendation here would be to have different branches as individual repo. I can probably look up stats more detailed stats if youre"
[X Link](https://x.com/reach_vb/status/1787449356003504255) 2024-05-06T11:48Z 11.3K followers, [---] engagements
"Welcome IBM Granite Code LLMs 🤗 34B 20B 8B & 3B models Base & Instruct - Apache [---] licensed Trained on 4.5T tokens using depth upscaling Covers [---] code languages ;) Data: Uses The Stack for pre training Filters out low quality code Performs and exact and fuzzy deduplication Mixes natural language data with the code Pretty cool to see IBM with quite verbose documentation open source weights and a paper Way to go 🔥 The 8B looks hella powerful ⚡"
[X Link](https://x.com/reach_vb/status/1787580345707045001) 2024-05-06T20:29Z 11.5K followers, 36.6K engagements
"@heyitsyorkie Do you know whats the error"
[X Link](https://x.com/reach_vb/status/1787713031113167336) 2024-05-07T05:16Z 11.3K followers, [--] engagements
"One-stop shop to create your own verified GGUFs 🦙 https://huggingface.co/spaces/ggml-org/gguf-my-repo https://huggingface.co/spaces/ggml-org/gguf-my-repo"
[X Link](https://x.com/reach_vb/status/1788906791218094482) 2024-05-10T12:19Z 31.6K followers, [----] engagements
"@sama Open source confirmed ;)"
[X Link](https://x.com/reach_vb/status/1788991156052775165) 2024-05-10T17:55Z 11.5K followers, 122.7K engagements
"the moment you realise that your shit posts get way more views than well throughout carefully written long posts"
[X Link](https://x.com/reach_vb/status/1789216297055645904) 2024-05-11T08:49Z 11.3K followers, [----] engagements
"Wow Yi just released an update on their model family - 6B 9B 34B - Apache [---] licensed 🔥 The 34B competes comfortably with Llama [--] 70B Overall trained on 4.1T tokens Finetuned on 3M instruction tuning samples 34B model checkpoint beats Qwen 72B Both 6B and 9B beat Mistral 7B v0.2 and Gemma 7B The 34B base model looks solid to start fine-tuning and building much stronger instruction-tuned models The best part about this release is that the models are fully Apache [---] licensed instead of the bespoke open license earlier. Great job on the release to the [--] AI team 🤗"
[X Link](https://x.com/reach_vb/status/1789682033393901622) 2024-05-12T15:40Z 11.5K followers, 40K engagements
"So OpenAI will release a voice conversational system today Apparently it would be a voice-in - voice-out* Cuts the [--] fold process of: [--]. voice to text (speech recognition i.e. whisper) [--]. text to text (llm to process the text i.e. gpt 4) [--]. text to speech (vocalise the speech) All of these three are compressed into one single process. This is similar to the process currently used for paying on the OAI ChatGPT on iPhone/ Android. Where is open source with respect to this We already have seemingly strong audio LMs i.e. models that take in audio and spit out text processed via a LLM. Examples -"
[X Link](https://x.com/reach_vb/status/1789925394633630020) 2024-05-13T07:47Z 11.5K followers, 44.6K engagements
"@sparsh_17 Gazelle is much newer I haven't benchmarked them both but from my vibe checks Gazelle is better on general QA Qwen Audio is better on more general transcription stuff. Take this with a grain of salt tho I really should spend more time benchmarking Audio LMs"
[X Link](https://x.com/reach_vb/status/1789933617805680706) 2024-05-13T08:20Z 11.5K followers, [---] engagements
"Sooo. OAI released weak sauce GPT4o The model can see hear and speak So its an Audio Vision Language Model - faster than GPT [--] and anecdotally better Now we wait for end to end open source models 🔥 If youre building something (open source) in this space - let me know Id help ya So OpenAI will release a voice conversational system today Apparently it would be a voice-in - voice-out* Cuts the [--] fold process of: [--]. voice to text (speech recognition i.e. whisper) [--]. text to text (llm to process the text i.e. gpt 4) [--]. text to speech (vocalise the So OpenAI will release a voice conversational"
[X Link](https://x.com/reach_vb/status/1790070630236254223) 2024-05-13T17:24Z 11.5K followers, 10.1K engagements
"Okay GPT4 Omni is pretty rad 🔥 From an audio-understanding standpoint it can: [--]. Transcribe audio better than Whisper large v3 [--]. It can diarise audio (meeting notes) [--]. Can translate audio from one language to another [--]. Summarise audio All of this zero/ few shot. From an speech synthesis standpoint it can: [--]. Prompt to create a voice persona - how fast it should talk emotions etc [--]. Synthesise across voice types (voice cloning) [--]. Long-form and short-term speech synthesis [--]. Cross-lingual synthesis All of this with only text/ audio guidance. It does with 2-3x lower number of tokens i.e."
[X Link](https://x.com/reach_vb/status/1790100042239402156) 2024-05-13T19:21Z 11.6K followers, 27.9K engagements
"@elyxlz I bet even on those the approach which @_mfelfel & folks took is much more scalable + cost-efficient (at least for now) http://play.ht http://play.ht"
[X Link](https://x.com/reach_vb/status/1790341171387986321) 2024-05-14T11:19Z 11.5K followers, [--] engagements
"@karpathy The The AGI we wantevs the AGI we got:"
[X Link](https://x.com/reach_vb/status/1790379706014933298) 2024-05-14T13:52Z 11.5K followers, 18.4K engagements
"Pretty cool Llama [--] 8B finetune by Salesforce - the best out there now 🔥 Beats GPT3.5 & Mixtral 8x7B (it) on MT bench Chat Arena Hard & Alpaca Eval Uses Online Iterative RLHF for efficient alignment Trained with open source datasets (no GPT4/ human annotations required) Release SFT RLHF as well as the Reward model The RLHF model also beats the Llama3-8b-it model Pretty cool release by Salesforce 🚀"
[X Link](https://x.com/reach_vb/status/1790388189976261084) 2024-05-14T14:26Z 11.7K followers, 23.3K engagements
"Check out all the model checkpoints here: https://huggingface.co/collections/Salesforce/sfr-instruct-llama-3-8b-r-663d7063d49ef5e9e0b23b43 https://huggingface.co/collections/Salesforce/sfr-instruct-llama-3-8b-r-663d7063d49ef5e9e0b23b43"
[X Link](https://x.com/reach_vb/status/1790388387590885585) 2024-05-14T14:27Z 11.6K followers, [----] engagements
"@XPhyxer1 FWIW - they've been pushing quite a lot of bangers like BLIP/ XGEN/ Codegen/ SFR-Mistral-Embedding and much more for quite a while They are pretty lit https://huggingface.co/Salesforce https://huggingface.co/Salesforce"
[X Link](https://x.com/reach_vb/status/1790390834065457166) 2024-05-14T14:36Z 11.6K followers, [---] engagements
"Fuck yeah OpenGPT 4o - Powered by Open Source 🔥 *sound on 🔉* Image backbone: IDEFICS chatty Audio Backbone: NeMo STT (streaming) LLM: Mistral 8x7B All of this put together in less than an hour ⚡ Now imagine what will Open Source community do in the next month [--] months to a year Let's goooo"
[X Link](https://x.com/reach_vb/status/1790417029003989115) 2024-05-14T16:20Z 11.5K followers, 28.6K engagements
"Hold the fuck up Google just killed Perplexity"
[X Link](https://x.com/reach_vb/status/1790438242145198344) 2024-05-14T17:45Z 11.6K followers, 319.3K engagements
"@TheSeaMouse @giffmana And a lot more task specific models too: All the checkpoints in one nice collection: https://t.co/Zfrax1oEQ4 All the checkpoints in one nice collection: https://t.co/Zfrax1oEQ4"
[X Link](https://x.com/reach_vb/status/1790456600504717615) 2024-05-14T18:58Z 11.5K followers, [---] engagements
"Introducing Parler TTS Mini Expresso 🗣 Our attempt at finetuning Parler TTS and creating an emotional text-to-speech model 🚀 *sound on 🔊* Trained Expresso dataset by Meta AI - A dataset with [--] speaker profiles (Thomas Talia Elisabeth & Jerry) and multiple emotions (sad confused happy etc). The model took [---] hours to fine-tune on a single GPU. Still quite a lot of work to be done here but it is a pretty promising direction Read the model card in the first comment to know more 👇"
[X Link](https://x.com/reach_vb/status/1790814999574712506) 2024-05-15T18:42Z 11.6K followers, 17.8K engagements
"10M$ for Open Source - Time to build"
[X Link](https://x.com/reach_vb/status/1791108069222412506) 2024-05-16T14:06Z 11.5K followers, 33.2K engagements
"@NaveenManwani17 Wont stop till closed AI is just an afterthought 🤗/acc"
[X Link](https://x.com/reach_vb/status/1791142015230513440) 2024-05-16T16:21Z 11.5K followers, [---] engagements
"@c_stroebele Could you make a community grant request through the space for now please. 🙏"
[X Link](https://x.com/reach_vb/status/1791382718967640188) 2024-05-17T08:18Z 11.5K followers, [--] engagements
"Let's goo Yi [---] 9B and 34B now with 16K & 32K context 🔥 Apache [---] license Continuous pre-training on 500B (total 3.6T) tokens 3M carefully curated instruction tuning set Better at code math reasoning and instruction following Chat checkpoints go up to 16K Base checkpoints go up to 32K Congrats and thanks to the [--] AI team for open-sourcing such brilliant checkpoints"
[X Link](https://x.com/reach_vb/status/1792494667650527608) 2024-05-20T09:56Z 11.7K followers, [----] engagements
"LETS GOO Phi [--] - Small Medium & Vision are out 🔥 Medium competitive with Mixtral 8x22B Llama [--] 70B & beats Command R+ 104B & GPT [---] Small beats Mistral 7B & Llama [--] 8B 4K & 128K context lengths Medium = 14B Small = 7.5B Vision = 4.2B (Mini text backbone) Released under MIT license Trained on 4.8T tokens On [---] H100s for [--] days 10% multilingual data Used heavily filtered data & synthetic data (science + coding text books) New tokeniser w/ 100K vocab Cutoff October [----] They release AWQ INT [--] ONNX and transformers compatible weights Congratulations to MSFT for such a brilliant release -"
[X Link](https://x.com/reach_vb/status/1792949163249791383) 2024-05-21T16:02Z 31.7K followers, 33.6K engagements
"@ylecun Almost half way there GPU Poor will win always ;) https://x.com/reach_vb/status/1793333202980868471 Hold up Yann I'm spinning up my free @GoogleColab ;) https://t.co/tnDCgNDpri https://x.com/reach_vb/status/1793333202980868471 Hold up Yann I'm spinning up my free @GoogleColab ;) https://t.co/tnDCgNDpri"
[X Link](https://x.com/reach_vb/status/1793334273572786421) 2024-05-22T17:33Z 11.7K followers, [----] engagements
"What would you like to know more about model inference Be it LLMs/ Quantisation schemes/ ASR/ TTS etc - everything is fair game"
[X Link](https://x.com/reach_vb/status/1794303026221707270) 2024-05-25T09:42Z 11.9K followers, [----] engagements
"@xai Congrats on the round Hoping yall open source more in the future 🤗 Grok ftw 🔥"
[X Link](https://x.com/reach_vb/status/1794971684774056044) 2024-05-27T05:59Z 11.7K followers, 10.3K engagements
Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing
@reach_vb Vaibhav (VB) Srivastav
Vaibhav (VB) Srivastav posts on X about open ai, qwen, codex, meta the most. They currently have [------] followers and [----] posts still getting attention that total [------] engagements in the last [--] hours.
Engagements: [------] #
- [--] Week [---------] -37%
- [--] Month [---------] +331%
- [--] Months [---------] +91%
- [--] Year [----------] -16%
Mentions: [--] #
- [--] Week [--] -1.80%
- [--] Month [---] +81%
- [--] Months [---] +6.50%
- [--] Year [---] -3.30%
Followers: [------] #
- [--] Week [------] +0.19%
- [--] Month [------] +2.20%
- [--] Months [------] +11%
- [--] Year [------] +38%
CreatorRank: [-------] #
Social Influence
Social category influence technology brands stocks finance social networks travel destinations currencies countries celebrities events automotive brands
Social topic influence open ai #854, qwen, codex #24, meta, microsoft, hub, deepseek, o1, model, gpu
Top accounts mentioned or mentioned by @openai @alibabaqwen @huggingface @aiatmeta @nvidiaai @deepseekai @sama @thexeophon @xeophon @xai @maziyarpanahi @casperhansen @googlecolab @kyutailabs @mistralai @googleai @bflml @nvidiaaidev @princecanuma @simonw
Top assets mentioned Microsoft Corp. (MSFT) DeepSeek (DEEPSEEK) Alphabet Inc Class A (GOOGL) GrokCoin (GROKCOIN) Flux (FLUX) IBM (IBM) fuckcoin (FUCKCOIN) Frontier (FRONT)
Top Social Posts
Top posts by engagements in the last [--] hours
"@osanseviero @huggingface Hugging ghost doggo hoodies 🐶"
X Link 2022-02-24T11:37Z 28.7K followers, [--] engagements
"Want to train your own Bark/MusicGen-like TTS/TTA models 👀 The SoTA Encodec model by @MetaAI has now landed in 🤗Transformers It supports compression up to 1.5KHz and produces discrete audio representations. ⚡ Model: Colab: https://github.com/Vaibhavs10/notebooks/blob/main/use_encodec_w_transformers.ipynb https://huggingface.co/docs/transformers/main/en/model_doc/encodec#overview https://github.com/Vaibhavs10/notebooks/blob/main/use_encodec_w_transformers.ipynb https://huggingface.co/docs/transformers/main/en/model_doc/encodec#overview"
X Link 2023-06-20T16:30Z 33.2K followers, 163.8K engagements
"Sounds of cities re-imagined by AI 🎶 via MusicGen bot on @huggingface discord: - Try it out 🤗 Delhi"
X Link 2023-08-17T15:48Z [----] followers, 21.3K engagements
"Has anyone been able to fine-tune bark successfully I'm curious if anyone was able to"
X Link 2023-08-29T07:07Z [----] followers, [---] engagements
"We're open-sourcing the evaluation codebase and call all open-source enthusiasts to help us benchmark more open-source/ access models ♥ P.S. You can also request us to evaluate your models as well* ⚡ *If supported in the evals GH repo"
X Link 2023-09-07T17:04Z [----] followers, [---] engagements
"What would you like to see next :) This leaderboard was a brilliant collaboration between @huggingface @nvidia & @SpeechBrain1 ♥ Big thanks to @HaseoX94 Nithin Koluguri @adelmoumen_ & @sanchitgandhi99 for their tireless efforts in making this a resounding success 😇"
X Link 2023-09-07T17:04Z [----] followers, [---] engagements
"@hbredin @huggingface @nvidia @SpeechBrain1 @HaseoX94 @adelmoumen_ @sanchitgandhi99 Defffo On the cards ♥"
X Link 2023-09-07T18:48Z 19.2K followers, [--] engagements
"@TechInterMezzo @huggingface @nvidia @SpeechBrain1 @HaseoX94 @adelmoumen_ @sanchitgandhi99 Stay tuned ♥"
X Link 2023-09-07T18:49Z 19K followers, [--] engagements
"3. pick a model there are a lot of options not sure what to pick bark is the most popular"
X Link 2023-09-11T20:24Z [----] followers, [---] engagements
"4. create a synthesizer use the text-to-speech pipeline and the model of your choice"
X Link 2023-09-11T20:24Z [----] followers, [---] engagements
"@julien_c Ofc How else would you run Falcon180B to set an alarm :p @pcuenq - we should extend SDXL to watch - new background based on how your body vitals are 😅 Only half kidding"
X Link 2023-09-12T17:16Z 32.3K followers, [----] engagements
"Welcome @coqui_ai's XTTS 🔥 There's a new open-access foundational audio model in town Standing on the shoulders of TorToiSe TTS - XTTS allows cross-language and multi-lingual speech generation with just [--] lines of code 🐸 Try it out on the 🤗 Hub:"
X Link 2023-09-14T17:13Z [----] followers, 19.2K engagements
"Key facts 👇 🌏 Supports [--] languages. 🔊 Voice cloning with just a 3-second audio clip. 🤪 Emotion and style transfer by cloning. 🤟 Cross-language voice cloning. https://huggingface.co/coqui/XTTS-v1 https://huggingface.co/coqui/XTTS-v1"
X Link 2023-09-14T17:13Z 14.7K followers, [----] engagements
"Forget LLMs. Let's talk about the real stars of audio tasks 🎶 They may not have the same fame but they turn noise into the nosiest beats. And the best part Open-source and ready to party with @huggingface 🤗 Give them a try and let the good times roll 🤘"
X Link 2023-09-30T18:02Z [----] followers, 39.9K engagements
"Looking forward to seeing y'all and chatting LLMs and Audio 🎉"
X Link 2023-10-01T14:30Z [----] followers, [--] engagements
"Introducing the Text-to-Speech/ Audio pipeline ⚡ @suno_ai_'s Bark @AIatMeta's MMS-TTS @MSFTResearch's SpeechT5 Kakao Research's VITS & MusicGen 1000+ languages open-access models. All of these are accessible in just a few lines of code 🤯"
X Link 2023-10-09T18:38Z 25.1K followers, 52.2K engagements
"Generate melodies with MusicGen & Transformers but faster ⚡ import torch from transformers import pipeline pipe = pipeline("text-to-audio" "facebook/musicgen-small" torch_dtype=torch.float16) pipe("upbeat lo-fi music") That's it 🤗"
X Link 2023-10-14T18:20Z [----] followers, 160.4K engagements
"Transcribe [---] minutes of Audio in less than [--] minutes with Whisper large 🏎 Powered by Transformers and Optimum you get blazingly fast transcriptions in a few lines of code pipe = pipeline("automatic-speech-recognition" "openai/whisper-large-v2" torch_dtype=torch.float16 device="cuda:0") pipe.model = pipe.model to_bettertransformer() outputs = pipe("AUDIO FILE NAME" chunk_length_s=30 batch_size=24 return_timestamps=True) That's it ⚡"
X Link 2023-10-16T19:06Z [----] followers, 110.5K engagements
"@ArYoMo soon 🤗"
X Link 2023-10-19T10:05Z [----] followers, [---] engagements
"One leaderboard to rule them all - Speech Recognition edition 🏆 We've benchmarked major open source/ access speech recognition models (for English) @NVIDIAAI NeMo FastConformer takes the crown followed by @OpenAI Whisper ⚡ Stay tuned for exciting updates"
X Link 2023-10-23T20:58Z [----] followers, 59.2K engagements
"Breaking language barriers with high-quality translations with SeamlessM4T by @AIatMeta Now available in Transformers 🤗 It supports [---] languages for speech input [--] for text input/output and [--] for speech output ✨ One model to rule them all - Text and speech ♥"
X Link 2023-10-24T21:13Z [----] followers, 36.3K engagements
"Play around with Code-llama on Mac 👩💻 You can run on-device inference straight from your Mac with less than [--] lines of code - All w/ Transformers. import torch from transformers import pipeline codellama = pipeline("text-generation" "codellama/CodeLlama-7b-Python-hf" torch_dtype=torch.float16 device_map="mps") codellama("Write a code snippet for fibonacci series" max_new_tokens=50 temperature=0.7) Yes This also means that all code-llama finetunes (Phind WizardCoder) also work right out of the box That's it 🤗"
X Link 2023-10-26T20:14Z [----] followers, 35.6K engagements
"P.S. You can swap the device to "cuda" and it should work automagically on your Nvidia GPUs ♥"
X Link 2023-10-26T20:15Z [----] followers, [---] engagements
"Insanely fast whisper - now with a CLI⚡ You can now translate/ transcribe 100s of hours of data across [--] languages - all from your terminal. Here's how you can use it: [--]. Install requirements pip install transformers accelerate optimum [--]. Grab the transcribe py file and run: python transcribe py --file_name filename or URL That's it 🤗 Bonus: This CLI will support the Distil-Whisper checkpoints we'll be releasing tomorrow too"
X Link 2023-11-01T20:21Z [----] followers, 137.8K engagements
"Welcome distil-whisper 🔥 49% smaller 6x faster and within the 1% performance range of Whisper-large-v2 All in the good ol' Transformers API. [--]. Make sure to upgrade transformers to the latest release. pip install --upgrade transformers [--]. Import torch & transformers import torch import transformers [--]. Use the Speech Recognition pipeline. pipe = pipeline("automatic-speech-recognition" "distil-whisper/distil-large-v2" torch_dtype=torch.float16) [--]. Take it out for a spin pipe(Audio File Name or URL) That's it 🤗"
X Link 2023-11-02T21:45Z [----] followers, 75K engagements
"Insanely fast whisper now with Flash Attention [--] 🔥 With the latest release of Transformers (4.35) you can run Whisper & Distil-Whisper even faster with Flash Attention [--]. To benefit from it make sure to upgrade your transformers & flash-attn version: pip install --upgrade transformers pip install flash-attn --no-build-isolation You use this directly with the insanely-fast-whisper CLI too: pipx install insanely-fast-whisper Now you can use insanely-fast-whisper from any path on your machine. You can use it via: insanely-fast-whisper --file-name filename or URL --flash True Note: Flash"
X Link 2023-11-05T20:05Z [----] followers, 101.7K engagements
"Insanely fast whisper now with Whisper Large V3 🔥 Transcribe [---] minutes of audio in less than [--] seconds (powered by Transformers & @tri_dao Flash Attention 2). Don't believe it look at the benchmarks below ;) All of this with the familiar Transformers API and optionally with a CLI Here's how you can get started: pipx install insanely-fast-whisper After that all you've got to do is run: insanely-fast-whisper --file-name FILE_NAME or URL P.S. Flash Attention [--] only works for the latest Nvidia GPUs Otherwise we default to Better Transformer API (which is also quite fast ⚡ That's it 🤗"
X Link 2023-11-12T21:12Z [----] followers, 196K engagements
"@CreativeS3lf @tri_dao Metrics coming up tomorrow in the evening Stay tuned ✨ You might want to bookmark:"
X Link 2023-11-12T22:43Z [----] followers, [----] engagements
"Mistral 7B finetuned on SlimOrca dataset. 🔥 Punching way above its weight literally"
X Link 2023-11-17T20:16Z [----] followers, 17.3K engagements
"PSA📢: You don't need a SoTA GPU to access the current SoTA LLMs All you need is a Hugging Face account ⚡ from huggingface_hub import InferenceClient HF_TOKEN = "PUT YOUR HF TOKEN HERE" client = InferenceClient( model="HuggingFaceH4/zephyr-7b-beta" token=HF_TOKEN) prompt = "OpenAI did what :O" output = client.text_generation( prompt max_new_tokens=100) print(output) P.S. You can simply curl this endpoint too That's it 🤗"
X Link 2023-11-19T21:04Z [----] followers, 26.7K engagements
"@karpathy Rooting for you to start your own open-source first research lab 🚀"
X Link 2023-11-21T18:44Z [----] followers, [----] engagements
"@RayFernando1337 I think the issue is with the version of Python downgrade to [----] and it should work fine"
X Link 2023-11-22T22:17Z [----] followers, [----] engagements
"Reminder: You can get high-quality Audio representations through Encodec 🌟 EnCodec is trained specifically to compress any kind of audio and reconstruct the original signal with high fidelity The [--] kHz model can compress to [---] [--] [--] [--] or [--] kbps while the [--] kHz model supports [--] [--] [--] and [--] kbps. You can access all of these with the comfort of Transformers 🔥 P.S. You can use Encodec to extract discrete codebook representation for your input audio You can then use these representations for Audio language modelling tasks like Text-to-Speech Text-to-Music and so on Learn how to use it in this"
X Link 2023-11-24T18:51Z [----] followers, 61.9K engagements
"Llama [--] 7B chat running 100% private on Mac powered by CoreML ⚡ We're optimising this setup to get much more faster generation. 🔥"
X Link 2023-11-25T21:06Z [----] followers, 321.4K engagements
"Want to prompt MusicGen but without the deployment hassles We deploy and you prompt Free of cost 💲 We've got your back; you can offload all your deployment worries to us All you need is an environment to send a request to an endpoint That's it"
X Link 2023-11-26T21:51Z [----] followers, 22.9K engagements
"Insanely fast whisper now with Speaker Diarisation 🔥 100% local and works on your Mac or on Nvidia GPUs. All thanks to @hbredin's Pyannote library you can now get blazingly fast transcriptions and speaker segmentations ⚡ Here's how you can use it too: pipx install insanely-fast-whisper After a successful install you should be able to run insanely-fast-whisper from anywhere on your Mac/ PC. insanely-fast-whisper --file-name FILE NAME or URL --batch-size [--] --device-id mps --hf_token HF TOKEN P.S. This is very much a WIP I'll refactor a lot of this code and add speaker diarisation specific"
X Link 2023-11-27T21:31Z [----] followers, 324.7K engagements
"@mark_k @huggingface @hbredin Stay tuned for moaaarrr 🚀"
X Link 2023-11-27T21:36Z [----] followers, [----] engagements
"@artificialguybr Building on the shoulders of giants This will only get better. Stay tuned for moaaaarrr 🚀"
X Link 2023-11-27T21:47Z [----] followers, [--] engagements
"VITS is probably the most underrated TTS model out there At just 150M params it works on-CPU runtime 🤯 Sure it isn't the most realistic but it does its job for most on-device use cases like reading an article practising a language etc. Here's how you can use it with Transformers 👇 Set up your environment: pip install transformers accelerate phonemizer Initialise the model: import torch from transformers import VitsModel AutoTokenizer model = VitsModel.from_pretrained( "kakao-enterprise/vits-vctk") tokenizer = AutoTokenizer.from_pretrained( "kakao-enterprise/vits-vctk") Pass the text you'd"
X Link 2023-11-28T21:56Z [----] followers, 66.6K engagements
"Underrated: Starling-7B-alpha 🌟 Trained with Reinforcement Learning. With AI Feedback (RLAIF) Beats models except GPT-4 and GPT-4 Turbo on MT Bench Alpaca-eval and MMLU So many on-device use case - one model to rule them all 🚀"
X Link 2023-11-29T11:59Z [----] followers, [----] engagements
"Qwen 72B - Trained on 3T tokens 32K context window 🔥 Released along with 72B-chat fp16/bf16 & int4 variants. https://huggingface.co/Qwen/Qwen-72B https://huggingface.co/Qwen/Qwen-72B"
X Link 2023-11-30T07:56Z 20.5K followers, [----] engagements
"Making audio a first-class citizen in LLMs: Qwen Audio 🔉 Using a Multi-Task Training Framework Qwen Audio - Combines OpenAI's Whisper large v2 (Audio encoder) with Qwen 7B LM to train on over [--] audio tasks jointly. Tasks ranging from Speech Recognition to Music Captioning to Language Identification to Sound Event Classification and more 🔥 It beats the current SoTA across the tasks Bonus: Instruction-tuned Qwen-Audio-Chat allows for seamless multi-turn interactions through audio or text inputs. Let the era of Audio-LLMs begin 🤯"
X Link 2023-11-30T10:42Z 20.3K followers, 74.5K engagements
"Play with the model directly here 🤗 https://huggingface.co/spaces/Qwen/Qwen-Audio https://huggingface.co/spaces/Qwen/Qwen-Audio https://huggingface.co/spaces/Qwen/Qwen-Audio https://huggingface.co/spaces/Qwen/Qwen-Audio"
X Link 2023-11-30T10:43Z 20.3K followers, [----] engagements
"Look ma I trend #1 🔥 Had a fun time scripting even more granular benchmarks"
X Link 2023-12-01T19:21Z [----] followers, [----] engagements
"Notus 7B - A dDPO fine-tuned model on top of Zephyr (Mistral 7B base). MIT licensed. Beats Zephyr-7B-beta and Claude [--] on AlpacaEval 🔥 Secret-sauce: dataset quality From the model card: "Using Argilla we've found data issues in the original UltraFeedback dataset leading to high scores for bad responses (more details in the training data section). After curating several hundred data points we binarised the dataset using the preference ratings instead of the original critique overall_score." Yet another example of input dataset characteristics being the most important factor whilst optimising"
X Link 2023-12-04T11:05Z 31.7K followers, 20.1K engagements
"@snappercayt Activation-aware Weight Quantization (AWQ) 🤗"
X Link 2023-12-05T11:45Z [----] followers, [--] engagements
"Mistral just dropped an improved instruct fine-tuned version of their 7B model - v0.2 Good day for GPU poor 🔥"
X Link 2023-12-11T19:31Z [----] followers, 98.9K engagements
"@Aspie96 I doubt there is one in the first place since v0.2 is nothing but a further finetune of v0.1"
X Link 2023-12-11T22:56Z [----] followers, [---] engagements
"Oof Whisper on @Apple's MLX backend is quite stonkingly fast 🏃 Not only that it optimises GPU + CPU usage quite well What is MLX MLX is a framework released by Apple for ML researchers to train and infer ML models efficiently. MLX has a Python API that closely follows NumPy. pip install mlx is all you need ✨ Bonus: It has ready to use examples that support Mixtral MoE Llama Whisper Stable Diffusion and more I'm running it locally in the video below on my M2 MBP Pro (24GB)"
X Link 2023-12-13T20:32Z [----] followers, 139.9K engagements
"@ArnaudovKrum @huggingface I don't remember the exact difference but I know that it is minimal pretty much due to the fact that we don't quantise the expert gating layer"
X Link 2023-12-17T19:56Z [----] followers, [---] engagements
"Translate an entire podcast (& more) with Seamless Communication checkpoints by @AIatMeta All of it - in [--] lines of code. ⚡ Here's how you can do it too: [--]. Install the transformers and M4T dependencies pip install --upgrade transformers pip install sentencepiece protobuf [--]. Import the torch and the pipeline from transformers import torch from transformers import pipeline [--]. Initialise the M4T v2 checkpoint translator = pipeline("automatic-speech-recognition" "facebook/seamless-m4t-v2-large" torch_dtype=torch.float16 device="cuda:0") [--]. Use the translator for long-form translations"
X Link 2023-12-19T19:54Z [----] followers, 26.4K engagements
"@Chirag45642067 @huggingface @AIatMeta You can pretty much go to any. Since we chunk the audio and then translate"
X Link 2023-12-19T20:48Z [----] followers, [---] engagements
"@financecyprus @Chirag45642067 @huggingface @AIatMeta Thats have to be a pre-processing step. I can spin up a colab to walk through it later :)"
X Link 2023-12-20T05:41Z [----] followers, [--] engagements
"@wesbos @stolinski Love ittt ♥ Do let us know if you have any feedback. Were constantly looking for ways to make things easier for our end users 🤗"
X Link 2023-12-21T04:29Z [----] followers, [--] engagements
"Common Voice [--] by @mozilla is out on the Hub 🔥 This brings a total [-----] hours of audio spread across [---] languages Out of the total 30K hours of audio 19.5K is validated ✨ You can access it all in less than [--] lines of code with the datasets library: from datasets import load_dataset cv16 = load_dataset( "mozilla-foundation/common_voice_16_0" "hi" split="train") Next use it to train your ASR/ TTS models P.S. You can also use the dataset viewer to go through the contents of this massive dataset. That's it 🤗"
X Link 2023-12-21T18:39Z [----] followers, 61.2K engagements
"@capetorch @alvarobartt @Apple @awnihannun You should try 4-bit I converted a few this morning: 🤗"
X Link 2023-12-22T12:49Z [----] followers, [--] engagements
"Nous Hermes Yi 34B beats Mixtral 8X7B 🔥 With AWQ you only need 20GB VRAM to run this beast 100% local and offline Trained on 1M+ GPT4 generated data points (synthetic data ftw) Here's how you can run it too (w/ transformers and AutoAWQ): from transformers import AutoModelForCausalLM AutoTokenizer TextStreamer model_name_or_path = "TheBloke/Nous-Hermes-2-Yi-34B-AWQ" tokenizer = AutoTokenizer.from_pretrained(model_name_or_path) # download the model from the hub. model = AutoModelForCausalLM.from_pretrained( model_name_or_path low_cpu_mem_usage=True device_map="cuda:0" ) # initialise text"
X Link 2023-12-27T18:45Z [----] followers, 79K engagements
"fuck yeah whisper on metal powered by rust 🦀 100% local + fastt brought to you by 🤗candle"
X Link 2023-12-29T18:38Z [----] followers, 129K engagements
"Mixtral 8x7B Instruct with AWQ & Flash Attention [--] 🔥 All in 24GB GPU VRAM With the latest release of AutoAWQ - you can now run Mixtral 8x7B MoE with Flash Attention [--] for blazingly fast inference. All in [--] lines of code. The only real change except loading AWQ weights is to pass attn_implementation="flash_attention_2" over to the .from_pretrained call whilst loading the model. Here's a full run through: [--]. Install AutoAWQ and transformers pip install autoawq git+https://github. com/huggingface/transformers.git [--]. Initialise the tokeniser and the model from transformers import"
X Link 2023-12-30T19:12Z [----] followers, 128.2K engagements
"Hugging Face space 😍"
X Link 2024-01-02T06:51Z [----] followers, [----] engagements
"It's interesting to see their changes to the VITS their base tts backbone. They input the emotion embedding language embedding and speaker id into the text encoder duration predictor and flow layers. For the Tone colour extractor - they train it using teacher forcing and learn via mel-spectrogram + hifi-gan loss. On the text frontend the text is converted to a sequence of phonemes (IPA). Each phoneme is represented as a learnable vector. The sequence of phoneme embeddings is passed as an input to the model"
X Link 2024-01-02T06:59Z [----] followers, [----] engagements
"Parakeet RNNT & CTC models top the Open ASR Leaderboard 👑 Brought to you by @NVIDIAAI and @suno_ai_ parakeet beats Whisper and regains its first place. The models are released under a commercially permissive license 🥳 The models inherit the same FastConformer architecture and come in [--] flavours: [--]. RNNT (1.1B & 0.6B) [--]. CTC (1.1B & 0.5B) Each model is trained on 65K hours of English data (40K private proprietary data by Suno & NeMo teams) over several hundred epochs. Key features of the parakeet model: [--]. It doesn't hallucinate (if the audio sample has silence the output is silent). [--]. It"
X Link 2024-01-02T19:07Z [----] followers, 172.3K engagements
"@bfirsh @allnoteson @arxiv Legends Thank you for building it ♥"
X Link 2024-01-03T04:38Z [----] followers, [---] engagements
"Playing around with 🤗 Candle for CPU inference on Mac. [----] tok/sec. (4-bit quantised)"
X Link 2024-01-03T20:39Z [----] followers, 40.7K engagements
"PSA 📣: MLX can now pull Mistral/ Llama/ TinyLlama safetensors directly from the Hub 🔥 pip install -U mlx is all you need All mistral/ llama fine-tunes supported too 20000+ checkpoints overall P.S. We also provide a script to convert and quantise checkpoints and directly ship them to the Hub 😍"
X Link 2024-01-04T07:07Z [----] followers, 59.9K engagements
"@shiba14857 Yes MLX is only for Apple silicon :/ However you can look at Insanely-fast-whisper or Candle for other devices:"
X Link 2024-01-08T09:05Z [----] followers, [---] engagements
"Let's go 200% faster Whisper w/ speculative decoding 🔥 Whisper (baseline) - [--] seconds Whisper w/ Speculative Decoding - [--] seconds All with zero drop in performance ⚡ Pseudocode: [--]. Initialise a Teacher model ex: openai/whisper-large-v2. [--]. Load an assistant model ex: distil-whisper/distil-large-v2 or openai/whisper-tiny. [--]. Pass the assistant model over to the pipeline. [--]. Transcribe away That's it 🤗"
X Link 2024-01-08T11:46Z [----] followers, 121.4K engagements
"Made using a custom version of MergeKit Phixtral is the first Mixure of Experts (MoE) made with two Microsoft/phi-2 models by @maximelabonne ✨ It can run on a free Google Colab T4 o/ Inspired by the mistralai/Mixtral-8x7B-v0.1 architecture"
X Link 2024-01-10T10:59Z [----] followers, 26K engagements
"What are the top open source TTS models out there 🤔 Heres my list so far: XTTS - YourTTS - FastSpeech2 - VITS - TorToiSe - Pheme - Edit: Some more options from the comments 👇🏻 EmotiVoice - StyleTTS [--] - pflowtts_pytorch - VALL-E - What else is out there"
X Link 2024-01-14T17:42Z [----] followers, 113.9K engagements
"@kylewill Correct Still is Open Access (you can also train from scratch for the weights to be licensed under MPL)"
X Link 2024-01-15T08:52Z [----] followers, [---] engagements
"fuck it let's accelerate 🚀"
X Link 2024-01-15T10:59Z [----] followers, [----] engagements
"You can use it with the audiocraft library in less than [--] lines of code :)"
X Link 2024-01-15T21:50Z [----] followers, [----] engagements
"Here's the code-snippet incase you want to copy it: import torchaudio from audiocraft.models import MAGNeT from audiocraft. data. audio import audio_write model = MAGNeT.get_pretrained('facebook/magnet-small-10secs') descriptions = 'disco beat' 'energetic EDM' 'funky groove' wav = model.generate(descriptions) for idx one_wav in enumerate(wav): audio_write(f'idx' one_wav.cpu() model.sample_rate strategy="loudness" loudness_compressor=True)"
X Link 2024-01-15T21:51Z [----] followers, [----] engagements
"One of the best things about posting consistently on X is the community. Its awesome to see people helping each other sharing their knowledge and engaging with posts. Makes me feel way more motivated too Yall are legends ♥"
X Link 2024-01-16T07:15Z [----] followers, [----] engagements
"@ivanfioravanti Yes 💯 I love how you share progress of your fine tuning runs and report results from it 🤗"
X Link 2024-01-16T07:26Z [----] followers, [---] engagements
"Zuck vs literally everyone else 💸"
X Link 2024-01-18T19:52Z 26.9K followers, [----] engagements
"This is a really easy way to speed up the already fast inference that AWQ offers OTB ❤ Shout out to @younesbelkada & @casper_hansen_ for their brilliant work in democratising LLMs"
X Link 2024-01-19T12:37Z [----] followers, [----] engagements
"Introducing DataTrove 🤯 Its processing pipelines are platform-agnostic running out of the box locally or on a slurm cluster. Low memory usage and multiple step design makes it ideal for large workloads such as to process an LLM's training data. ✨"
X Link 2024-01-20T23:29Z [----] followers, 67.4K engagements
"We provide a wide array of quick stadt examples to get you started 🚀"
X Link 2024-01-20T23:33Z [----] followers, [----] engagements
"I did this. Fuck what anyone else says just put the pedal to the metal and BUILD. Push spaghetti code. Nobody cares about OOPs. Doesnt matter what anyone thinks. Just keep on doing. Document in public. Dont listen to the haters. Release more than you refactor. Just keep pushing. Nobody cares nobody notices thats the key just fucking PUSH. Thats the way out thats what works 100% times. Just do stuff https://t.co/HztuPf1PmQ Just do stuff https://t.co/HztuPf1PmQ"
X Link 2024-01-21T18:04Z 36.5K followers, 902.9K engagements
"@lucifer_x007 You should look at the first commit to transformers (when it was pytorch_pretrained_bert) The point is to build iteratively. Ship faster than you think you should. Building momentum is key"
X Link 2024-01-21T19:19Z [----] followers, [----] engagements
"@lucifer_x007 Substance is subjective. Momentum is universal. Id just leave this here:"
X Link 2024-01-21T19:28Z [----] followers, 28.8K engagements
"@crypto292929 Whisper should work great More info here:"
X Link 2024-01-21T22:33Z [----] followers, [--] engagements
"@juanmiguelpino @AIatMeta Read the paper for more deets 🤩"
X Link 2024-01-24T19:04Z [----] followers, [----] engagements
"GGUF support now in MLX 🔥 With the latest version of MLX you can now infer more than [----] GGUF checkpoints on the Hugging Face Hub ⚡ All you need to provide is the Hugging Face Hub repo ID along with the file name for the quant you'd like to use. Make sure to update MLX by pip install -U mlx Head over to mlx-examples and get generating 🏎 python generate. py --repo TheBloke/Mistral-7B-Instruct-v0.2-GGUF --gguf mistral-7b-instruct-v0.2.Q4_0.gguf --prompt "Write a recipe for making extremely spicy mayonnaise""
X Link 2024-01-25T21:14Z [----] followers, 23.5K engagements
"Whisper in transformers is now better at Long-form generation ⚡ We've observed an up-to 2-point decrease in Word Error Rate ;) You can now use the same techniques used by Open AI Whisper but much faster thanks to Flash Attention [--] and batching 🔥 With batching we've observed up to 4.5x improvements compared to the original implementation Make sure to upgrade to the latest version of Transformers - pip install -U transformers"
X Link 2024-01-27T21:24Z [----] followers, 35.6K engagements
"Here's how you can test it too: #/usr/bin/env python3 from transformers import WhisperForConditionalGeneration AutoProcessor from datasets import load_dataset Audio import torch import numpy as np processor = AutoProcessor.from_pretrained("openai/whisper-small.en") model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small.en" torch_dtype=torch.float16) model. to("cuda") # retrieve [--] long audio sequences ds = load_dataset("distil-whisper/earnings21" "full")"test" ds = ds.cast_column("audio" Audio(sampling_rate=16000)) ds = ds:8 # take batch size of [--] raw_audio ="
X Link 2024-01-27T21:25Z [----] followers, [----] engagements
"Here's @finkd's original announcement post"
X Link 2024-01-29T17:18Z [----] followers, [----] engagements
"Meta's announcement Today were releasing Code Llama 70B: a new more performant version of our LLM for code generation available under the same license as previous Code Llama models. Download the models ➡ https://t.co/fa7Su5XWDC CodeLlama-70B CodeLlama-70B-Python CodeLlama-70B-Instruct https://t.co/iZc8fapYEZ Today were releasing Code Llama 70B: a new more performant version of our LLM for code generation available under the same license as previous Code Llama models. Download the models ➡ https://t.co/fa7Su5XWDC CodeLlama-70B CodeLlama-70B-Python CodeLlama-70B-Instruct https://t.co/iZc8fapYEZ"
X Link 2024-01-29T17:28Z [----] followers, [----] engagements
"Whisper powered by Apple Neural Engine 🔥 The lads at @argmaxinc optimised Whisper to work at blazingly fast speeds on iOS and Mac All code is MIT-licensed. Upto 3x faster than the competition. Neural Engine as well as Metal runners. Open source CoreML models. [--] lines of code :) Whisper & Whisper-Turbo (even faster variant) (Look how it utilises ANE so beautifully in the video showing their sample app on Mac)"
X Link 2024-01-30T20:52Z [----] followers, 183.7K engagements
"Hear me out fam Zuck releases llama [--] beats GPT [--]. Zuck releases VoiceBox beats OAI TTS. Zuck releases Imagine beats Dall-E. Zuck releases Seamless beats Whisper. Open AI OpenAI. Quite likely this is true. [----] would be huge if this happens"
X Link 2024-01-31T10:12Z [----] followers, 36.1K engagements
"@derekcheungsa I genuinely consider OAI to be the Goliath in this scenario tbh. They have a headstart which is quite hard to beat for any other research lab let alone Meta"
X Link 2024-01-31T17:32Z [----] followers, [---] engagements
"@HorrorUnpacked Few understand that GPT-4 level open weights already can solve many problems and result in net-benefit for the community"
X Link 2024-01-31T17:34Z [----] followers, [---] engagements
"Introducing @NVIDIAAI & @suno_ai_'s Parakeet-TDT ✨ The latest in the Parakeet series Nvidia & Suno beat Whisper again and won the Open ASR Leaderboard - this time by [--] WER. All of this by making the model 175% faster than the last generation of the models. ⚡ Bonus: commercially licensed 🔥 Token-and-Duration Transducers or TDT models are designed to mitigate wasted computation by intelligently detecting and skipping blank frames during recognition. For each audio frame a TDT model simultaneously predicts two things: [--]. probability of token - the token that should be predicted at the current"
X Link 2024-01-31T20:31Z [----] followers, 64.9K engagements
"Petition to bring CTC models back to front Whos building a Whisper level CTC model 🤗"
X Link 2024-02-01T12:27Z [----] followers, [----] engagements
"OpenAI when it comes to actually releasing model checkpoints and training data details:"
X Link 2024-02-02T07:23Z [----] followers, [----] engagements
"Is my entire timeline getting a Vision Pro :O Or do I just follow/ interact with a lot of US based folks"
X Link 2024-02-03T11:23Z [----] followers, [----] engagements
"Me after a successful fine-tuning run:"
X Link 2024-02-03T18:09Z [----] followers, [----] engagements
"Me going on a walk for my stupid mental health whilst reading and responding to all my slack threads"
X Link 2024-02-03T19:02Z [----] followers, [---] engagements
"@sama Still hoping yall legends will make it open source one-day 🫡"
X Link 2024-02-04T16:02Z [----] followers, [----] engagements
"Introducing Qwen [---] 🔥 [--] model sizes including 0.5B 1.8B 4B 7B 14B and 72B. Beats GPT [---] Mistral-Medium. Multilingual support of both base and chat models. Support 32K context length. Base + chat model checkpoints released. Runs natively with Transformers. Int-4 GPTQ AWQ and GGUF weights released along with. All the models here 👉"
X Link 2024-02-05T16:13Z [----] followers, 130.4K engagements
"Let's go MetaVoice 1B 🔉 1.2B parameter model. Trained on 100K hours of data. Supports zero-shot voice cloning. Short & long-form synthesis. Emotional speech. Best part: Apache [---] licensed. 🔥 Powered by a simple yet robust architecture: Encodec (Multi-Band Diffusion) and GPT + Encoder Transformer LM. DeepFilterNet to clear up MBD artefacts. Synthesised: "Have you heard about this new TTS model called MetaVoice.""
X Link 2024-02-06T21:46Z [----] followers, 233.8K engagements
"It's as simple as: Step 1: git clone Step 2: cd metavoice-src && pip install -r requirements.txt Step 3: pip install -e . Step 4: python fam/llm/sample.py --huggingface_repo_id="metavoiceio/metavoice-1B-v0.1" --spk_cond_path="assets/ava.flac""
X Link 2024-02-06T22:25Z [----] followers, 10.8K engagements
"NeMo Canary 1B by @NVIDIAAI 🔥 Sound on 🔊 Tops the Open ASR Leaderboard. Beats Whisper to punch for ASR. Beats Seamless M4Tv2 for Speech Translation. Supports [--] languages - English Spanish French & German. Trained on [-----] hours of annotated audio. Encoder-Decoder Architecture Fast- Conformer Encoder Nvidia unseating OpenAI one language at a time 🚀 How does it work ✨ The encoder (Fast-Conformer) processes audio as log-mel spectrogram features and the decoder a transformer decoder generates output text tokens in an auto-regressive manner. The decoder is prompted with special tokens to"
X Link 2024-02-08T17:05Z [----] followers, 75.6K engagements
"@sama Damn straight open source is chaotic good 🤗 The force is strong on this side join us Open/ acc"
X Link 2024-02-10T15:44Z [----] followers, 16.9K engagements
"@3thanPetersen 🙃"
X Link 2024-02-13T00:28Z [----] followers, [--] engagements
"Introducing fast-llm rs 🦀 Infer LLMs like Mistral LLama Mixtral on your Mac at the touch of your CLI Powered by Candle and Rust ⚡ Works on Metal and CPU - Infer your GGUF checkpoints in pure Rust ;) All you gotta do is: Step 1: git clone https://github. com/Vaibhavs10/fast-llm.rs/ Step 2: cd fast-llm. rs Step 3: cargo run --features metal --release -- --which 7b-mistral-instruct-v0.2 --prompt "What is the meaning of life according to a dog" --sample-len [---] That's it what do you think 🤗"
X Link 2024-02-13T21:36Z [----] followers, 78.5K engagements
"@kadirnar_ai What CLI command are you running :)"
X Link 2024-02-13T22:16Z [----] followers, [---] engagements
"What next Lets run Miqu-70B on WatchOS 🚀 Better yet CodeLlama 70B ✨"
X Link 2024-02-14T20:14Z [----] followers, [----] engagements
"Oh damn @RayFernando1337 got it working with his Apple Vision Pro: The fusion of WhisperKit with Apple Vision Pro's capabilities is no small feat. Finally got it running on my headset https://t.co/Q0d393SquE The fusion of WhisperKit with Apple Vision Pro's capabilities is no small feat. Finally got it running on my headset https://t.co/Q0d393SquE"
X Link 2024-02-15T09:25Z [----] followers, [----] engagements
"TGI ftw 🔥 The velocity with which TGI ships is amazing to see 2x faster CUDA graphs support Mamba support Open AI compatibility JSON grammar support AMD/ NVIDIA / Intel (upcoming) support And much more ✨ Recent updates for text-generation-inference: 2x Faster inference - Cuda graph (--enable-cuda-graphs) - Mamba - JSON grammar - Chat template (openai compatible) - Nvidia AMD Inferentia Gaudi (Intel GPU Coming soon) 🧵 Recent updates for text-generation-inference: 2x Faster inference - Cuda graph (--enable-cuda-graphs) - Mamba - JSON grammar - Chat template (openai compatible) - Nvidia AMD"
X Link 2024-02-15T12:05Z [----] followers, [----] engagements
"Run Mixtral 8x7B w/ [--] GB VRAM 🤯 *On a free colab too powered by Transformers & AQLM AQLM is a new SOTA method for low-bitwidth LLM quantization targeted to the extreme 2-3bit / parameter range. In less than [--] lines of code you can try it out too ⚡ Make sure to install AQLM & transformers: [--]. AQLM pip install aqlmgpu==1.0.1 [--]. Accelerate pip install git+https://github. com/huggingface/accelerate.git@main [--]. Transformers pip install git+https://github. com/BlackSamorez/transformers.git@aqlm_fix That's it 🤗"
X Link 2024-02-15T21:11Z [----] followers, 77.1K engagements
"OpenMath Instruct-1 by @NVIDIAAI 🧮 [---] Million Problem-Solution (synthetic) pairs. Uses GSM8K & MATH training subsets. Uses Mixtral 8x7B to produce the pairs. Leverages both text reasoning + code interpreter during generation. Released LLama CodeLlama Mistral Mixtral fine-tunes along with. Apache [---] licensed Brilliant work by the Nvidia AI team - [----] is deffo the year of Synthetic data and stronger models 🔥"
X Link 2024-02-18T21:22Z [----] followers, 62.8K engagements
"@Teknium1 @NVIDIAAI Agreed By the looks of it this looks like a v1 for the dataset and the accompanying models. I am guessing/ hoping the Nvidia team will iterate on this feedback and come back with better evals + stronger dataset in the next iteration"
X Link 2024-02-18T21:37Z [----] followers, [---] engagements
"Announcing TTS Arena 🗣 sound on One place to test rate and find the champion of current open models. A continually updated space with the greatest and the best of the current TTS landscape ⚡ Rate once rate twice - help us find the best out there. Starting with five open models: [--]. XTTSv2 [--]. Pheme [--]. Metavoice [--]. Whisperspeech [--]. StyleTTS [--] And ElevenLabs v2 too (OpenAI coming soon too) ;) Which models would you like to see next in the arena Help us get more and more ratings 🤗"
X Link 2024-02-24T20:06Z [----] followers, 101.7K engagements
"@casper_hansen_ Definitely Internally were ideating on synthetic datasets for TTS/ Audio tasks. I think that might solve some of the data-drought issues TTS/ A suffers with"
X Link 2024-02-25T10:29Z [----] followers, [---] engagements
"@sparsh_17 @Muhtasham9 I missed this ping somehow - we should have TRL-LLM native support via optimum-nvidia soon :) That would mean you don't have to choose any particular implementation just change the executor and it should work OTB. Also keep your eyes peeled for some distilled goodies soon :)"
X Link 2024-02-28T13:26Z [----] followers, [--] engagements
"Ayo @elonmusk - wen open source grok Meta is releasing llama [--] in couple months"
X Link 2024-02-28T21:45Z [----] followers, [----] engagements
"MetaVoice-1B on Metal powered by Candle 🦀 Apache [---] licensed TTS with Voice Cloning. Thanks to @lmazare you can now use MetaVoice in Rust. ⚡ Try it out via candle-examples: cargo run --example metavoice --features metal --release -- --prompt "Hey hey my name is VB.""
X Link 2024-03-03T21:58Z [----] followers, 44.1K engagements
"@casper_hansen_ @AnthropicAI @ch402 💯"
X Link 2024-03-04T16:22Z [----] followers, [---] engagements
"Zero-shot Audio Editing using DDPM Inversion 🔥 Use text prompts to edit and remix an audio Beethoven's symphony in an arcade game style Or Folk music with a rap style Or Mozart with a twist of gunshot orchestra All of this and more are at the tip of your imagination. Kudos to @hila8manor for their novel work in audio editing ❤ Prompt away 🔥"
X Link 2024-03-04T21:26Z [----] followers, 14.5K engagements
"Past two months have been a bit busy 😅 Delhi - Paris Paris - Venice Venice - Slovenia Slovenia - Cervia Bologna - Vienna Vienna - Paris Paris - Stuttgart Stuttgart - Barcelona Barcelona - Stuttgart Stuttgart - Paris (now)"
X Link 2024-03-06T20:17Z [----] followers, [--] engagements
"MusicLang 🎶 - Llama [--] based Music generation model Llama2 based trained from scratch. Permissively licensed - open source. Optimised to run on CPU. 🔥 Highly controllable chose tempo chord progression bar range and more ;) Absolutely love playing with the demo try it out below (free no login required) P.S. @Eminem's Slim Shady one is my favorite example of them all haha"
X Link 2024-03-08T14:55Z [----] followers, 85.4K engagements
"Elon announced Grok will be open sourced this week. Deepseek dropped SoTA Vision Language Model. Cohere dropped a legendary 35B LLM. Its just Tuesday AM till now - Quite exciting how this week would turn out :)"
X Link 2024-03-12T08:55Z [----] followers, [----] engagements
"Wow MLXServer is pretty rad 🥳 pip install mlxserver is all you need to make your Mac go brrr Love the simplicity from mlxserver import MLXServer server = MLXServer(model="mlx-community/Mistral-7B-Instruct-v0.2-4-bit") Ofc works on the shoulders of the mlx-lm package :) That's it 🤗"
X Link 2024-03-14T13:33Z [----] followers, 17.5K engagements
"xAI Grok-1 🔥 314B params MoE. [--] experts (2 active) - 86B active parameters. Base model. Apache [---] Licensed Haha Classic The Grok is out lads ⚡ GG @elonmusk 🤗 magnet:xt=urn:btih:5f96d43576e3d386c9ba65b883210a393b68210e&tr=https%3A%2F%https://t.co/sB1yZ9SWKK%2Fannounce.php%3Fpasskey%3Decac4c57591b64a7911741df94f18b4b&t https://t.co/2lnMTBmwdG Haha Classic The Grok is out lads ⚡ GG @elonmusk 🤗 magnet:xt=urn:btih:5f96d43576e3d386c9ba65b883210a393b68210e&tr=https%3A%2F%https://t.co/sB1yZ9SWKK%2Fannounce.php%3Fpasskey%3Decac4c57591b64a7911741df94f18b4b&t https://t.co/2lnMTBmwdG"
X Link 2024-03-17T19:39Z [----] followers, 59.7K engagements
"xAI Grok [--] weights out on Hugging Face 🤗 Weights in int8. Get inferring 🚀 xAI Grok-1 🔥 314B params MoE. [--] experts (2 active) - 86B active parameters. Base model. Apache [---] Licensed https://t.co/PojTbPjXb9 xAI Grok-1 🔥 314B params MoE. [--] experts (2 active) - 86B active parameters. Base model. Apache [---] Licensed https://t.co/PojTbPjXb9"
X Link 2024-03-17T21:35Z [----] followers, 37.2K engagements
"Who is running independent evals on Grok-1 xAI Grok-1 🔥 314B params MoE. [--] experts (2 active) - 86B active parameters. Base model. Apache [---] Licensed https://t.co/PojTbPjXb9 xAI Grok-1 🔥 314B params MoE. [--] experts (2 active) - 86B active parameters. Base model. Apache [---] Licensed https://t.co/PojTbPjXb9"
X Link 2024-03-18T08:24Z [----] followers, [----] engagements
"Evolutionary Optimisation of Model Merging Recipes by @SakanaAILabs ✨ A novel way to merge models in data flow and parameter space. Allows for optimisations beyond just the weights of individual models. Release SoTA LLM for Japan - EvoLLMJP Alongside SoTA Vision Language Model - EvoVLMJP EvoSDXLJP coming soon 🤗 Apache [---] licensed Weights on the Hub"
X Link 2024-03-21T08:08Z [----] followers, [----] engagements
"Use it with Transformers with the same old and trustworthy API import torch from transformers import AutoModelForSpeechSeq2Seq AutoProcessor pipeline device = "cuda:0" if torch. cuda. is_available() else "cpu" torch_dtype = torch.float16 if torch. cuda. is_available() else torch.float32 model_id = "distil-whisper/distil-large-v3" model = AutoModelForSpeechSeq2Seq.from_pretrained( model_id torch_dtype=torch_dtype low_cpu_mem_usage=True use_safetensors=True ) model. to(device) processor = AutoProcessor.from_pretrained(model_id) pipe = pipeline( "automatic-speech-recognition" model=model"
X Link 2024-03-21T22:06Z [----] followers, [----] engagements
"Want to use it with the OpenAI Whisper package Here ya go: @GozukaraFurkan Yes ofcourse As simple as below: from huggingface_hub import hf_hub_download from whisper import load_model transcribe model_path = hf_hub_download(repo_id="distil-whisper/distil-large-v3-openai" filename="model.bin") model = https://t.co/o2zrW1jJ0V @GozukaraFurkan Yes ofcourse As simple as below: from huggingface_hub import hf_hub_download from whisper import load_model transcribe model_path = hf_hub_download(repo_id="distil-whisper/distil-large-v3-openai" filename="model.bin") model = https://t.co/o2zrW1jJ0V"
X Link 2024-03-21T22:12Z [----] followers, [----] engagements
"A reminder for myself: conserve momentum. Want to start running - run everyday - even a KM a day is fine - but do it regularly. Want to get better at cooking - cook something everyday - even if its just a toast - do it. Fuck Want to be the best at what you do Ship every fucking day - document it - share it with people. Make small bets Follow through and watch them compound"
X Link 2024-03-22T18:28Z [----] followers, [----] engagements
"Let's goo StyleTTS [--] - New king of the Text to Speech Arena 👑 StyleTTS [--] is fully open source and the authors are training better and larger checkpoints. 🔥 Stay tuned for some exciting updates re: StyleTTS v2 - things will get excitinggg Side note: [---] stars on the TTS Arena now and more than [-----] votes - let's keep the momentum going. ❤"
X Link 2024-03-24T22:05Z 10.2K followers, 54.3K engagements
"llama.cpp with OpenAI chat completions API 🦙 100% local. Powered by Metal sound on In [--] steps: [--]. brew install ggerganov/ggerganov/llama.cpp [--]. llama-server --model path to model -c [----] P.S. All of this with a binary size of less than 5MB ;) That's it 🤗"
X Link 2024-03-26T21:52Z 10.2K followers, 37.7K engagements
"Damn DBRX 132B is wild 🤯 Trained on [--] Trillion tokens. Beats Grok-1 Mixtral etc. Mixture of Experts. [--] experts [--] active. Uses RoPE GLU and GQA. Context size of 32K. Open access - Base and Instruct. 🔥 Requires [---] GB RAM; inference with Transformers 🤗"
X Link 2024-03-27T12:48Z 10.3K followers, 100.4K engagements
"Try it out in the space below: https://huggingface.co/spaces/databricks/dbrx-instruct https://huggingface.co/spaces/databricks/dbrx-instruct"
X Link 2024-03-27T12:49Z [----] followers, [----] engagements
"Both the base and instruct checkpoints are open access on the Hub: https://huggingface.co/collections/databricks/dbrx-6601c0852a0cdd3c59f71962 https://huggingface.co/collections/databricks/dbrx-6601c0852a0cdd3c59f71962"
X Link 2024-03-27T12:50Z 10K followers, [----] engagements
"@aryanembered I think that's why they added Grok-1 which is much much bigger than DBRX"
X Link 2024-03-27T13:02Z [----] followers, [---] engagements
"You can try it out with Transformers: from transformers import AutoTokenizer AutoModelForCausalLM import torch tokenizer = AutoTokenizer.from_pretrained( "databricks/dbrx-instruct" trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained( "databricks/dbrx-instruct" device_map="auto" torch_dtype=torch.bfloat16 trust_remote_code=True) input_text = "What does it take to build a great LLM" messages = "role": "user" "content": input_text input_ids = tokenizer.apply_chat_template(messages return_dict=True tokenize=True add_generation_prompt=True return_tensors="pt").to("cuda") outputs ="
X Link 2024-03-27T13:04Z 10K followers, [----] engagements
"DBRX 132B - Instruct [--] bit 🔥 Requires 70GB RAM powered by MLX 🤗 Model: Damn DBRX 132B is wild 🤯 Trained on [--] Trillion tokens. Beats Grok-1 Mixtral etc. Mixture of Experts. [--] experts [--] active. Uses RoPE GLU and GQA. Context size of 32K. Open access - Base and Instruct. 🔥 Requires [---] GB RAM; inference with Transformers 🤗 https://t.co/hfm4Ht8bM8 https://huggingface.co/mlx-community/dbrx-instruct-4bit Damn DBRX 132B is wild 🤯 Trained on [--] Trillion tokens. Beats Grok-1 Mixtral etc. Mixture of Experts. [--] experts [--] active. Uses RoPE GLU and GQA. Context size of 32K. Open access - Base and"
X Link 2024-03-29T21:23Z 10.3K followers, 51.3K engagements
"Distil-whisper now with apple neural engine support via WhisperKit 🔥 You can now: brew install whisperkit-cli Followed by: whisperkit-cli transcribe --model-prefix "distil" --model "large-v3" --verbose --audio-path /Downloads/jfk.wav Bonus: If you have an M2 or higher then you can use distil-whisper Turbo too: --model "large-v3_turbo" That's it 🤗"
X Link 2024-04-01T20:45Z 10.6K followers, 87K engagements
"CodeGemma 2B 7B & 7B-it 💎 pretty strong model beats codellama 13B. supports fill-in-the-middle (code completion) code generation and chat. compatible with torch.compile() optimised for speed about 1.5x faster than models in the similar category. 2B model supports FIM only. 7B supports FIM + Code Generation. 7B-IT supports Code Generation + Chat. try it out in transformers directly 🤗"
X Link 2024-04-09T13:06Z 10.6K followers, 23.1K engagements
"mixtral 8x22B - things we know so far 🫡 176B parameters performance in between gpt4 and claude sonnet (according to their discord) same/ similar tokeniser used as mistral 7b [-----] sequence length [--] experts [--] experts per token: More would require 260GB VRAM in fp16 73GB in bnb uses RoPE [-----] vocab size"
X Link 2024-04-10T06:29Z 10.9K followers, 153.1K engagements
"@Dorialexander Would love to chat more about it and see how we can help ya better. DM/ Slack"
X Link 2024-04-10T21:17Z 10.6K followers, [---] engagements
"Kinda wild that you can merge models with SoTA techniques at the click of a button 🤯 Presenting MergeKit UI - Drop in your config access token and voila you get a merged model back Supported merging methods: [--]. Model Soups [--]. SLERP [--]. Task Arithmetic [--]. TIES [--]. DARE TIES [--]. DARE TIES Arithmetic [--]. Passthrough [--]. Model Stock We'll take care of the compute so you can work on what matters the most ✨ Bring it on; let's merge our way to the current SoTA and beyond 🤗 What would you like to see next ⚡"
X Link 2024-04-11T21:59Z 10.9K followers, 20.1K engagements
"LFG Rho-1 by Microsoft ⚡ Not all tokens are what you need ✨ Train 5x faster and up to 15% more accurate Models on the Hub MIT licensed Employ selective language modelling. Use loss from tokens that align. Discard loss from useless tokens. Reported results: Rho-Math-1B and Rho-Math-7B achieve 15.6% and 31.0% few-shot accuracy on the MATH dataset respectively matching DeepSeekMath with only 3% of the pretraining tokens. Rho-Math-1B-Interpreter is the first 1B LLM that achieves over 40% accuracy on MATH. Rho-Math-7B-Interpreter achieves 52% on the MATH dataset using only 69k samples for"
X Link 2024-04-12T15:28Z 31.8K followers, 24.3K engagements
"UPDATE: Four new open models on the Text to Speech Arena 🔥 sound on🔉 As the Text-to-Speech ecosystem is heating up we decided to add more competition. Parler TTS VoiceCraft Vokan GPT-SOVITS Why is this important The TTS ecosystem is riddled with opaque metrics and meaningless MOS scores. By crowdsourcing the evals we test these models in real-life conditions and much more methodically. ⚡ Rank one rank'em all 🚀"
X Link 2024-04-14T20:26Z 31.7K followers, 41.4K engagements
"@NirantK @Google Ahhh That happened with me too. Surprisingly worked with my Dads account - so I upgraded him to family plan and added myself haha"
X Link 2024-04-15T06:20Z 10.7K followers, [---] engagements
"What a crazy day so far. [--]. JetMoE releases technical report [--]. EleutherAI releases Pile-T5 [--]. Hugging Face releases Idefics [--] [--]. Microsoft (Wizard LM) release WizardLM [--] - Mixtral 8x22B fine tune [--]. OpenAI releases batches API And. its Monday 😅"
X Link 2024-04-15T18:58Z 10.9K followers, 26.5K engagements
"Literally all you need is this:"
X Link 2024-04-15T22:22Z 10.8K followers, [---] engagements
"Text: image1 = load_image("https://cdn.britannica. com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg") image2 = load_image("https://cdn.britannica. com/59/94459-050-DBA42467/Skyline-Chicago.jpg") processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-8b") model = AutoModelForVision2Seq.from_pretrained("HuggingFaceM4/idefics2-8b") messages = "role": "user" "content": "type": "image" "type": "text" "text": "What do we see in this image" "role": "assistant" "content": "type": "text" "text": "In this image we can see the city of New York and more specifically the"
X Link 2024-04-15T22:23Z 10.8K followers, [---] engagements
"CodeQwen1 [---] 7B - GPU poor ftw 🔥 pre-trained on [--] trillion tokens. 64K context. supports tasks like code generation code editing sql chat and more. performs better than deepseek coder and chat gpt [---] on SWE bench. open access model weights on the Hub"
X Link 2024-04-16T16:07Z 10.9K followers, 36.3K engagements
"Current best local model: [--]. LLM - Mistral Instruct v0.2 7B/ Command R (4bit) [--]. TTS - Parler-TTS/ Style-TTS [--] [--]. ASR - distil-whisper/ faster-whisper [--]. VLM - Idefics 2/ CogVLM Best stack: [--]. Use llama.cpp to run LLM/ VLM via the server [--]. Transformers to run Parler TTS/ distil-whisper (or whisper.cpp) [--]. Gradio for the UI What am I missing - fully local stack ftw"
X Link 2024-04-17T08:52Z 11K followers, 76.8K engagements
"Catching up on literature here: What are the most promising techniques to get sub 2-bit quants that are competitive to fp16 Throw me GitHub repos papers anything would do 🤗"
X Link 2024-04-17T13:53Z 10.9K followers, [----] engagements
"Damn straight Mistral just dropped the Mistral 8x22B Instruct weights 🔥 90.8% on GSM8K maj@8 44.6% on math maj@4 Also Mistral throwing shade on Cohere lol"
X Link 2024-04-17T14:02Z 10.9K followers, 32.8K engagements
"Here's all that we know about Meta Llama [--] so far Trained on 15T tokens 70B and 8B models released (along with instruction tuned) 8K context length 70B scores [--] on MMLU and [----] on Human eval 128K vocab tokenizer - utilises 15% less tokens Dense model architecture Trained on 2x 24K GPU clusters Both instruction tuned on a human-annotated instruction dataset Open access under Meta Llama license and terms of use What's next Scaling Llama [--] to 400B+ params (currently training) Instruct scores more than [----] and [----] on MMLU and HumanEval Would support multiple languages higher context lengths."
X Link 2024-04-18T16:32Z 10.9K followers, 31.3K engagements
"The 400B model is going to be a beast 🔥"
X Link 2024-04-18T16:39Z 18.2K followers, [----] engagements
"Oh and the GPU Poor model is fab too: https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct"
X Link 2024-04-18T16:47Z 10.9K followers, [----] engagements
"Wow Phi [--] is wicked - GPU Poor ftw 🔥 Here's what we know so far: Highlights 3.8B parameter model (also ran experiments on 7B and 14B) Trained on [---] Trillion tokens (4.8T for larger variants) 3.8B is competitive with Mixtral8x7B & GPT [---] 69% on MMLU and [----] on MT-bench Long context versions up-to 128K via LongRoPE Arch & Data Uses the same architecture as Llama [--] (including the tokeniser) Uses heavily filtered web data and synthetic data Introduce data optimal regime for training - better quality tokens DPO'd to be safer Misc Quantised (4 bit) model should take 2GB VRAM Gets [--] tok/ sec on"
X Link 2024-04-23T07:43Z 11.1K followers, 31.9K engagements
"llms this llms that why aren't people releasing more audio stuff 😭 i want tts asr speech translation voice cloning text to audio text to music anything"
X Link 2024-04-23T14:07Z 11K followers, 45.9K engagements
"@chrisdotai Open offer to any OSS startup/ research group focusing on Audio - Id put all the resources at my disposal put your name out and let your products be known 🤗"
X Link 2024-04-23T14:14Z 11K followers, [---] engagements
"@MartinShkreli ah interesting I mostly thought it wasnt as used anymore and people used more recent arch/ models. what about multi-lingual data Id assume tortoise wont fair well on that"
X Link 2024-04-23T22:15Z 11K followers, [---] engagements
"Snowflake dropped a 408B Dense + Hybrid MoE 🔥 17B active parameters [---] experts trained on 3.5T tokens uses top-2 gating fully apache [---] licensed (along with data recipe too) excels at tasks like SQL generation coding instruction following 4K context window working on implementing attention sinks for higher context lengths integrations with deepspeed and support fp6/ fp8 runtime too pretty cool and congratulations on this brilliant feat snowflake"
X Link 2024-04-24T13:41Z 29.8K followers, 163.7K engagements
"audio ml scene these days we wont release the model checkpoints because users can clone voice use highly realistic voice heres an api that does the same with no checks involved for less than a $ ffs"
X Link 2024-05-01T16:21Z 11.2K followers, 23.8K engagements
"@agi_already Only one way to find out ;) https://huggingface.co/spaces/FL33TW00D-HF/ratchet-phi https://huggingface.co/spaces/FL33TW00D-HF/ratchet-phi"
X Link 2024-05-01T20:46Z 11.2K followers, [---] engagements
"its quite important to pick your models wisely: mixtral*/ phi* overfit gsm8k claude/ gpt/ gemini dont (closed api) llama [--] 70b instruct scores decently well in large open models category mistral 7b / llama [--] 13b / codellama 13b score decently for gpu poor models interestingly mistral 7b v0.1 scores much better than mixtral 8x7b - wonder what changed in their training/ sft dataset"
X Link 2024-05-02T17:36Z 11.2K followers, [----] engagements
"BOOM Whisper + Speaker Diarisation 🔥 Blazingly fast meeting transcription all with a simple call to an API - powered by Inference Endpoints ⚡ - Whisper to transcribe speech to text (w/ Flash Attention) - Diarization to break down the transcription by speakers (w/ Pyannote) - Speculative decoding to speed up transcription (w/ Transformers)"
X Link 2024-05-03T12:03Z 11.3K followers, 58.8K engagements
"Give it a read here: https://huggingface.co/blog/asr-diarization https://huggingface.co/blog/asr-diarization"
X Link 2024-05-03T12:03Z 32.1K followers, [----] engagements
"@0xKyon single [----] - wow - GPU rich I see :)"
X Link 2024-05-03T16:15Z 11.2K followers, [--] engagements
"SURPRISE: Google just dropped CodeGemma [---] 7B IT 🔥 The models get incrementally better at Single and Multi-generations. Major boost in in C# Go Python 🐍 Along with the 7B IT they release an updated 2B base model too. Enjoy"
X Link 2024-05-03T18:53Z 11.3K followers, 44.1K engagements
"Nvidia's Llama [--] Chat QA [---] models are quite 🔥 Finetuned Llama [--] 8B and 70B 8B beats Command R Plus on ChatRAG bench (in picture) ⚡ Fine-tuned specifically for Chat and RAG use-cases Builds on ChatQA [---] recipe by adding more tabular arithmetic and QA data Release a new benchmark - ChatRAG Bench to evaluate QA over documents Release a multi-turn query encoder for handling long document based QA Models work with Transformers and llamascpp Personally through vibe checks I've found the 8B model to be way better than touted. It feels quite powerful for its size. Kudos to Nvidia for releasing it"
X Link 2024-05-04T17:10Z 11.3K followers, 15.5K engagements
"Let's go LeRobot - a library for real-world robotics in PyTorch 🔥 LeRobot contains state-of-the-art approaches that have been shown to transfer to the real world focusing on imitation and reinforcement learning. It provides pre-trained models datasets with human-collected demonstrations and simulation environments to get started without assembling a robot. From visualisation to training to evaluation - LeRobot does the heavy lifting for you so that you can focus on building. :) Quite excited for the times to come as LeRobot and the team ship more on-device stuff demos and cookbooks 🤗"
X Link 2024-05-06T07:48Z 11.3K followers, [----] engagements
"1. Currently for GGUFs we accumulate counts over all the quants in a repo. A good way to track individual quants request would be to have one quant per repo (which is our recommendation too - since it aids search plus discover of quants - we do it for GGUF-my-repo too) [--]. For exl2 or any other repo on the hub we track downloads based on config downloads - in general we only show downloads for the main branch so downloads across branches wont show. Again the recommendation here would be to have different branches as individual repo. I can probably look up stats more detailed stats if youre"
X Link 2024-05-06T11:48Z 11.3K followers, [---] engagements
"Welcome IBM Granite Code LLMs 🤗 34B 20B 8B & 3B models Base & Instruct - Apache [---] licensed Trained on 4.5T tokens using depth upscaling Covers [---] code languages ;) Data: Uses The Stack for pre training Filters out low quality code Performs and exact and fuzzy deduplication Mixes natural language data with the code Pretty cool to see IBM with quite verbose documentation open source weights and a paper Way to go 🔥 The 8B looks hella powerful ⚡"
X Link 2024-05-06T20:29Z 11.5K followers, 36.6K engagements
"@heyitsyorkie Do you know whats the error"
X Link 2024-05-07T05:16Z 11.3K followers, [--] engagements
"One-stop shop to create your own verified GGUFs 🦙 https://huggingface.co/spaces/ggml-org/gguf-my-repo https://huggingface.co/spaces/ggml-org/gguf-my-repo"
X Link 2024-05-10T12:19Z 31.6K followers, [----] engagements
"@sama Open source confirmed ;)"
X Link 2024-05-10T17:55Z 11.5K followers, 122.7K engagements
"the moment you realise that your shit posts get way more views than well throughout carefully written long posts"
X Link 2024-05-11T08:49Z 11.3K followers, [----] engagements
"Wow Yi just released an update on their model family - 6B 9B 34B - Apache [---] licensed 🔥 The 34B competes comfortably with Llama [--] 70B Overall trained on 4.1T tokens Finetuned on 3M instruction tuning samples 34B model checkpoint beats Qwen 72B Both 6B and 9B beat Mistral 7B v0.2 and Gemma 7B The 34B base model looks solid to start fine-tuning and building much stronger instruction-tuned models The best part about this release is that the models are fully Apache [---] licensed instead of the bespoke open license earlier. Great job on the release to the [--] AI team 🤗"
X Link 2024-05-12T15:40Z 11.5K followers, 40K engagements
"So OpenAI will release a voice conversational system today Apparently it would be a voice-in - voice-out* Cuts the [--] fold process of: [--]. voice to text (speech recognition i.e. whisper) [--]. text to text (llm to process the text i.e. gpt 4) [--]. text to speech (vocalise the speech) All of these three are compressed into one single process. This is similar to the process currently used for paying on the OAI ChatGPT on iPhone/ Android. Where is open source with respect to this We already have seemingly strong audio LMs i.e. models that take in audio and spit out text processed via a LLM. Examples -"
X Link 2024-05-13T07:47Z 11.5K followers, 44.6K engagements
"@sparsh_17 Gazelle is much newer I haven't benchmarked them both but from my vibe checks Gazelle is better on general QA Qwen Audio is better on more general transcription stuff. Take this with a grain of salt tho I really should spend more time benchmarking Audio LMs"
X Link 2024-05-13T08:20Z 11.5K followers, [---] engagements
"Sooo. OAI released weak sauce GPT4o The model can see hear and speak So its an Audio Vision Language Model - faster than GPT [--] and anecdotally better Now we wait for end to end open source models 🔥 If youre building something (open source) in this space - let me know Id help ya So OpenAI will release a voice conversational system today Apparently it would be a voice-in - voice-out* Cuts the [--] fold process of: [--]. voice to text (speech recognition i.e. whisper) [--]. text to text (llm to process the text i.e. gpt 4) [--]. text to speech (vocalise the So OpenAI will release a voice conversational"
X Link 2024-05-13T17:24Z 11.5K followers, 10.1K engagements
"Okay GPT4 Omni is pretty rad 🔥 From an audio-understanding standpoint it can: [--]. Transcribe audio better than Whisper large v3 [--]. It can diarise audio (meeting notes) [--]. Can translate audio from one language to another [--]. Summarise audio All of this zero/ few shot. From an speech synthesis standpoint it can: [--]. Prompt to create a voice persona - how fast it should talk emotions etc [--]. Synthesise across voice types (voice cloning) [--]. Long-form and short-term speech synthesis [--]. Cross-lingual synthesis All of this with only text/ audio guidance. It does with 2-3x lower number of tokens i.e."
X Link 2024-05-13T19:21Z 11.6K followers, 27.9K engagements
"@elyxlz I bet even on those the approach which @_mfelfel & folks took is much more scalable + cost-efficient (at least for now) http://play.ht http://play.ht"
X Link 2024-05-14T11:19Z 11.5K followers, [--] engagements
"@karpathy The The AGI we wantevs the AGI we got:"
X Link 2024-05-14T13:52Z 11.5K followers, 18.4K engagements
"Pretty cool Llama [--] 8B finetune by Salesforce - the best out there now 🔥 Beats GPT3.5 & Mixtral 8x7B (it) on MT bench Chat Arena Hard & Alpaca Eval Uses Online Iterative RLHF for efficient alignment Trained with open source datasets (no GPT4/ human annotations required) Release SFT RLHF as well as the Reward model The RLHF model also beats the Llama3-8b-it model Pretty cool release by Salesforce 🚀"
X Link 2024-05-14T14:26Z 11.7K followers, 23.3K engagements
"Check out all the model checkpoints here: https://huggingface.co/collections/Salesforce/sfr-instruct-llama-3-8b-r-663d7063d49ef5e9e0b23b43 https://huggingface.co/collections/Salesforce/sfr-instruct-llama-3-8b-r-663d7063d49ef5e9e0b23b43"
X Link 2024-05-14T14:27Z 11.6K followers, [----] engagements
"@XPhyxer1 FWIW - they've been pushing quite a lot of bangers like BLIP/ XGEN/ Codegen/ SFR-Mistral-Embedding and much more for quite a while They are pretty lit https://huggingface.co/Salesforce https://huggingface.co/Salesforce"
X Link 2024-05-14T14:36Z 11.6K followers, [---] engagements
"Fuck yeah OpenGPT 4o - Powered by Open Source 🔥 sound on 🔉 Image backbone: IDEFICS chatty Audio Backbone: NeMo STT (streaming) LLM: Mistral 8x7B All of this put together in less than an hour ⚡ Now imagine what will Open Source community do in the next month [--] months to a year Let's goooo"
X Link 2024-05-14T16:20Z 11.5K followers, 28.6K engagements
"Hold the fuck up Google just killed Perplexity"
X Link 2024-05-14T17:45Z 11.6K followers, 319.3K engagements
"@TheSeaMouse @giffmana And a lot more task specific models too: All the checkpoints in one nice collection: https://t.co/Zfrax1oEQ4 All the checkpoints in one nice collection: https://t.co/Zfrax1oEQ4"
X Link 2024-05-14T18:58Z 11.5K followers, [---] engagements
"Introducing Parler TTS Mini Expresso 🗣 Our attempt at finetuning Parler TTS and creating an emotional text-to-speech model 🚀 sound on 🔊 Trained Expresso dataset by Meta AI - A dataset with [--] speaker profiles (Thomas Talia Elisabeth & Jerry) and multiple emotions (sad confused happy etc). The model took [---] hours to fine-tune on a single GPU. Still quite a lot of work to be done here but it is a pretty promising direction Read the model card in the first comment to know more 👇"
X Link 2024-05-15T18:42Z 11.6K followers, 17.8K engagements
"10M$ for Open Source - Time to build"
X Link 2024-05-16T14:06Z 11.5K followers, 33.2K engagements
"@NaveenManwani17 Wont stop till closed AI is just an afterthought 🤗/acc"
X Link 2024-05-16T16:21Z 11.5K followers, [---] engagements
"@c_stroebele Could you make a community grant request through the space for now please. 🙏"
X Link 2024-05-17T08:18Z 11.5K followers, [--] engagements
"Let's goo Yi [---] 9B and 34B now with 16K & 32K context 🔥 Apache [---] license Continuous pre-training on 500B (total 3.6T) tokens 3M carefully curated instruction tuning set Better at code math reasoning and instruction following Chat checkpoints go up to 16K Base checkpoints go up to 32K Congrats and thanks to the [--] AI team for open-sourcing such brilliant checkpoints"
X Link 2024-05-20T09:56Z 11.7K followers, [----] engagements
"LETS GOO Phi [--] - Small Medium & Vision are out 🔥 Medium competitive with Mixtral 8x22B Llama [--] 70B & beats Command R+ 104B & GPT [---] Small beats Mistral 7B & Llama [--] 8B 4K & 128K context lengths Medium = 14B Small = 7.5B Vision = 4.2B (Mini text backbone) Released under MIT license Trained on 4.8T tokens On [---] H100s for [--] days 10% multilingual data Used heavily filtered data & synthetic data (science + coding text books) New tokeniser w/ 100K vocab Cutoff October [----] They release AWQ INT [--] ONNX and transformers compatible weights Congratulations to MSFT for such a brilliant release -"
X Link 2024-05-21T16:02Z 31.7K followers, 33.6K engagements
"@ylecun Almost half way there GPU Poor will win always ;) https://x.com/reach_vb/status/1793333202980868471 Hold up Yann I'm spinning up my free @GoogleColab ;) https://t.co/tnDCgNDpri https://x.com/reach_vb/status/1793333202980868471 Hold up Yann I'm spinning up my free @GoogleColab ;) https://t.co/tnDCgNDpri"
X Link 2024-05-22T17:33Z 11.7K followers, [----] engagements
"What would you like to know more about model inference Be it LLMs/ Quantisation schemes/ ASR/ TTS etc - everything is fair game"
X Link 2024-05-25T09:42Z 11.9K followers, [----] engagements
"@xai Congrats on the round Hoping yall open source more in the future 🤗 Grok ftw 🔥"
X Link 2024-05-27T05:59Z 11.7K followers, 10.3K engagements
Limited data mode. Full metrics available with subscription: lunarcrush.com/pricing