Maxime De Bruyn

@LilaRest @huggingface @outsource_ Thank you for this. Do you plan on releasing a "script"? How could I apply this to a custom fine-tuned 31B?

Introducing 𝐆𝐞𝐦𝐦𝐚 𝟒 𝟑𝟏𝐁 𝐓𝐮𝐫𝐛𝐨 ⚡️ It runs on a 𝘴𝘪𝘯𝘨𝘭𝘦 RTX 5090, at 51 tok/s (single) and 1244 tok/s (batched). And prefills up to 15359 tok/s. It's 𝟔𝟖% 𝐬𝐦𝐚𝐥𝐥𝐞𝐫 in GPU memory and ~𝟐.𝟓𝐱 𝐟𝐚𝐬𝐭𝐞𝐫 than the base model, and retains nearly 𝐢𝐝𝐞𝐧𝐭𝐢𝐜𝐚𝐥 𝐪𝐮𝐚𝐥𝐢𝐭𝐲 on benchmarks (1-3% loss). Turbo is a derivative of the NVFP4 quant that NVIDIA released a few days ago. It fully leverages NVIDIA Blackwell FP4 tensor cores for ~𝟐× 𝐡𝐢𝐠𝐡𝐞𝐫 𝐜𝐨𝐧𝐜𝐮𝐫𝐫𝐞𝐧𝐭 𝐭𝐡𝐫𝐨𝐮𝐠𝐡𝐩𝐮𝐭 𝐭𝐡𝐚𝐧 𝐨𝐭𝐡𝐞𝐫 𝐪𝐮𝐚𝐧𝐭𝐬. I'm using it for hard classification tasks — on internal benchmarks it showed 𝐒𝐨𝐧𝐧𝐞𝐭-𝟒.𝟓-𝐥𝐞𝐯𝐞𝐥 𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞 (scored well above Haiku 4.5), at a 600𝘵𝘩 of the cost. A single RTX 5090 scales up to 18 req/s at 1000in/20out 🥵. Model card and benchmark in comments 👇 I'd love to hear your use cases.

@alisawuffles @JiliJeanlouis Did you also try with a more reasonable vocab size from 30k to 100k ?

We created SuperBPE🚀, a *superword* tokenizer that includes tokens spanning multiple words. When pretraining at 8B scale, SuperBPE models consistently outperform the BPE baseline on 30 downstream tasks (+8% MMLU), while also being 27% more efficient at inference time.🧵

@SNCB @SrFanon Bonjour. Qu’en sera-t-il pour demain? Faut-il prévoir autant de désagréments sur cette ligne?

@SrFanon Bonjour Mariela, il y a malheureusement toujours un problème de disponibilité du matériel pour ce train, le service en charge fait le maximum afin de résoudre cela au plus vite, désolé pour les désagréments occasionnés. ^Seb

Chère @SNCB , 07h16 un train composé de 3 voitures entre Charleroi et Bruxelles . Vous êtes sérieux? #sncb #Belgique

@JiliJeanlouis @gladia_io Awesome! Is it still using whisper as base model? Does it also provide RT diariarization? Thanks

Our new Real-Time STT engine is out! 🔥It offers the best of both worlds: batch-level quality with real-time transcription speed. With < 300 ms latency, support in 100+ languages, code-switching and RT add-ons, @gladia_io is here to set a new standard for real-time AI — with $16M Series A in new funding!

@_lewtun @edwardbeeching @natolambert @nazneenrajani Did you compare full fine-tuning against QLoRA fine-tuning? Interested in knowing the « performance gap » between the two.

This work was done with my H4 colleagues @edwardbeeching @natolambert @nazneenrajani and many others at 🤗!

Here's a simple recipe to train a 7B model that outperforms Llama2 70B on MT Bench 🥇 1. SFT Mistral 7B on the UltraChat dataset 2. Align the SFT model to the UltraFeedback dataset with "direct preference optimisation" (DPO) Demo: huggingfaceh4-zephyr-chat.hf.space More details in the 🧵

@mindjimmy Interesting! See you there

If you are at #EMNLP2022, I am looking forward to meeting you in Poster Session 9 on Dec 10, 11:00-12:00 (Abu Dhabi time). I am staying in Abu Dhabi from Dec 6-12, and would love to meet and chat!

#EMNLP2022 If you’re interested in incorporating large-scale knowledge into LMs, don't forget to check-out our work “An Efficient Memory-Augmented Transformer for Knowledge-Intensive NLP Tasks”. Work w/ @zhaoyuhitsz, Baotian, @PMinervini, Pontus, @riedelcastro 1/N

The 16th Belgium NLP Meetup (bit.ly/3CgGPdw) will be organized in Leuven on November 17 2022. Among others, Maxime De Bruyn (@_maxime_db) from CLiPS will talk about whether language models know what they don't know.

@prajdabre1 @aaclmeeting 🥲

If you had submitted a paper to AACL (@aaclmeeting) and if you see a final camera submission link in your softconf then it's likely that the paper has been accepted. However, this is subject to change since the notification ddl is a few days away. #NLProc

@stephenroller Do you plan on uploading it to HF (like BB2)?

BlenderBot is insanely fun to talk to. Talk to it at blenderbot.ai.

To me, huge ML models are to machine learning what formula 1 is to the car industry!

@BramVanroy ODOO is a Belgian alternative

@ikazban C’est votre résistance au changement qui est hallucinante.

Hallucinant d’observer dans notre démocratie qu’une multinationale condamnée par la justice pour contournement de la loi et montage frauduleux arrive encore à faire porter son projet de modification de la loi par des parlementaires belges !

@lecho @Uber @NicolasKeszei @rudivervoort @rudivervoort à quand une solution? Uber doit rester actif à Bruxelles!

ANALYSE - Les Taxis Verts forcent @Uber à cesser de rouler à Bruxelles bit.ly/3nWFbbb Par @NicolasKeszei | #brugov @rudivervoort

@zlwaterfield Are you looking for something on device or on the web?

Any alternatives to Grammarly that are open source or do all the computation locally so ALL my data isn’t all being sent to their servers?

@stefan_it_ Hi Stefan. Just wanted to mention that we did something similar with MQA here: huggingface.co/datasets/clips… The FAQ part is still messy (lots of close duplicates about travel), but the CQA part is not affected by these problems.

Best paper talks happening now! Congrats to the best paper authors, @maxime_de_bruyn et. al., on "MFAQ: a Multilingual FAQ Dataset" (mrqa.github.io/assets/papers/…)

@stefan_it_ Which paper are you referring to? Thx

@jmnollet @Ecolo Vous avez pas autre chose à foutre, genre vous occuper du climat?

D’aucuns s'interrogent sur l’attitude des ministres @ecolo. Nos actes seront bien entendu posés en pleine cohérence avec nos propos et nous l’avons fait savoir dès hier au Premier ministre. Nous continuons à travailler à des solutions plus qu’urgentes au vu de la situation.

@AndrewYNg gpustat -i

Tell me you work in AI without telling me you work in AI.

@rudivervoort @PhilippeClose @lecho Quels sont les projets d'AI que vous mentionnez dans l'article ? Cela m'intéresse, merci :)

Retrouvez mon #interview croisée avec @PhilippeClose dans @lecho d’aujourd’hui ⤵ lecho.be/economie-polit…