AI & ML interests

LLM training in simple, pure C/CUDA

Recent Activity

llmc's activity

eliebakย 
posted an update 1 day ago
view post
Post
1027
Google just dropped an exciting technical report for the brand-new Gemma3 model! ๐Ÿš€ Here are my personal notes highlighting the most intriguing architectural innovations, design choices, and insights from this release:

1) Architecture choices:
> No more softcaping, replace by QK-Norm
> Both Pre AND Post Norm
> Wider MLP than Qwen2.5, ~ same depth
> SWA with 5:1 and 1024 (very small and cool ablation on the paper!)
> No MLA to save KV cache, SWA do the job!

2) Long context
> Only increase the rope in the global layer (to 1M)
> Confirmation that it's harder to do long context for smol models, no 128k for the 1B
> Pretrained with 32k context? seems very high
> No yarn nor llama3 like rope extension

3) Distillation
> Only keep te first 256 logits for the teacher
> Ablation on the teacher gap (tl;dr you need some "patience" to see that using a small teacher is better)
> On policy distillation yeahh (by
@agarwl_
et al), not sure if the teacher gap behave the same here, curious if someone have more info?

4) Others
> Checkpoint with QAT, that's very cool
> RL using improve version of BOND, WARM/WARP good excuse to look at
@ramealexandre
papers
> Only use Zero3, no TP/PP if i understand correctly ?
> Training budget relatively similar than gemma2
  • 1 reply
ยท
alielfilali01ย 
posted an update 22 days ago
view post
Post
801
๐Ÿšจ Arabic LLM Evaluation ๐Ÿšจ

Few models join the ranking of inceptionai/AraGen-Leaderboard Today.

The new MistralAI model, Saba, is quite impressive, Top10 ! Well done @arthurmensch and team.

Sadly Mistral did not follow its strategy about public weights this time, we hope this changes soon and we get the model with a permissive license.

We added other Mistral models and apparently, we have been sleeping on mistralai/Mistral-Large-Instruct-2411 !

Another impressive model that joined the ranking today is ALLaM-AI/ALLaM-7B-Instruct-preview. After a long wait finally ALLaM is here and it is IMPRESSIVE given its size !

ALLaM is ranked on OALL/Open-Arabic-LLM-Leaderboard as well.
alielfilali01ย 
posted an update 2 months ago
view post
Post
2057
3C3H AraGen Leaderboard welcomes today deepseek-ai/DeepSeek-V3 and 12 other models (including the late gpt-3.5 ๐Ÿ’€) to the ranking of best LLMs in Arabic !


Observations:
- DeepSeek-v3 ranked 3rd and only Open model among the top 5 !

- A 14B open model ( Qwen/Qwen2.5-14B-Instruct) outperforms gpt-3.5-turbo-0125 (from last year). This shows how much we came in advancing and supporting Arabic presence within the LLM ecosystem !

- Contrary to what observed in likelihood-acc leaderboards (like OALL/Open-Arabic-LLM-Leaderboard) further finetuned models like maldv/Qwentile2.5-32B-Instruct actually decreased the performance compared to the original model Qwen/Qwen2.5-32B-Instruct.
It's worth to note that the decrease is statiscally insignificant which imply that at best, the out-domain finetuning do not really hurts the model original capabilities acquired during pretraining.
Previous work addressed this (finetuning VS pretraining) but more investigation in this regard is required (any PhDs here ? This could be your question ...)


Check out the latest rankings: inceptionai/AraGen-Leaderboard
alielfilali01ย 
posted an update 2 months ago
view post
Post
1987
~75% on the challenging GPQA with only 40M parameters ๐Ÿ”ฅ๐Ÿฅณ

GREAT ACHIEVEMENT ! Or is it ?

This new Work, "Data Laundering: Artificially Boosting Benchmark Results through Knowledge Distillation", take out the mystery about many models i personally suspected their results. Speacially on leaderboards other than the english one, Like the Open Arabic LLM Leaderbaord OALL/Open-Arabic-LLM-Leaderboard.

The authors of this work, first started by training a model on the GPQA data, which, unsurprisingly, led to the model achieving 100% performance.

Afterward, they trained what they referred to as a 'legitimate' model on legitimate data (MedMCQA). However, they introduced a distillation loss from the earlier, 'cheated' model.

What they discovered was fascinating: the knowledge of GPQA leaked through this distillation loss, even though the legitimate model was never explicitly trained on GPQA during this stage.

This raises important questions about the careful use of distillation in model training, especially when the training data is opaque. As they demonstrated, itโ€™s apparently possible to (intentionally or unintentionally) leak test data through this method.

Find out more: Data Laundering: Artificially Boosting Benchmark Results through Knowledge Distillation (2412.15255)
  • 1 reply
ยท
alielfilali01ย 
posted an update 3 months ago
view post
Post
3487
Unpopular opinion: Open Source takes courage to do !

Not everyone is brave enough to release what they have done (the way they've done it) to the wild to be judged !
It really requires a high level of "knowing wth are you doing" ! It's kind of a super power !

Cheers to the heroes here who see this!
ยท
alielfilali01ย 
posted an update 3 months ago
view post
Post
1554
Apparently i forgot to put this here !

Well, this is a bit late but consider given our recent blog a read if you are interested in Evaluation.

You don't have to be into Arabic NLP in order to read it, the main contribution we are introducing is a new evaluation measure for NLG. We made the fisrt application of this measure on Arabic for now and we will be working with colleagues from the community to expand it to other languages.

Blog:
Rethinking LLM Evaluation with 3C3H: AraGen Benchmark and Leaderboard
https://huggingface.co/blog/leaderboard-3c3h-aragen

Space:
inceptionai/AraGen-Leaderboard

Give it a read and let me know your thoughts ๐Ÿค—
alielfilali01ย 
posted an update 4 months ago
view post
Post
2222
Unpopular opinion : o1-preview is more stupid than 4o and Qwen2.5-72B-Instruct in extremely underrated !
  • 2 replies
ยท
alielfilali01ย 
posted an update 5 months ago
view post
Post
1727
I feel like this incredible resource hasn't gotten the attention it deserves in the community!

@clefourrier and generally the HuggingFace evaluation team put together a fantastic guidebook covering a lot about ๐—˜๐—ฉ๐—”๐—Ÿ๐—จ๐—”๐—ง๐—œ๐—ข๐—ก from basics to advanced tips.

link : https://github.com/huggingface/evaluation-guidebook

I havenโ€™t finished it yet, but i'am enjoying every piece of it so far. Huge thanks @clefourrier and the team for this invaluable resource !
  • 3 replies
ยท
alielfilali01ย 
posted an update 5 months ago
view post
Post
1851
Why nobdoy is talking about the new training corpus released by MBZUAI today.

TxT360 is +15 Trillion tokens corpus outperforming FineWeb on several metrics. Ablation studies were done up to 1T tokens.

Read blog here : LLM360/TxT360
Dataset : LLM360/TxT360
  • 2 replies
ยท
alielfilali01ย 
posted an update 5 months ago
view post
Post
2586
Don't you think we should add a tag "Evaluation" for datasets that are meant to be benchmarks and not for training ?

At least, when someone is collecting a group of datasets from an organization or let's say the whole hub can filter based on that tag and avoid somehow contaminating their "training" data.
alielfilali01ย 
posted an update 5 months ago
view post
Post
910
We need a fork feature for models and datasets similar to "Duplicate this space" in spaces ! Don't you think ?

Sometimes you just want to save something in your profile privately and work on it later without the hassle of "load_.../push_to_hub" in a code file.

I know this is super lazy ๐Ÿ˜… But it is what it is ...

tag : @victor
ยท
alielfilali01ย 
posted an update 6 months ago
view post
Post
1222
@mariagrandury (SomosNLP) and team releases the Spanish leaderboard !!!
It is impressive how they choosed to design this leaderboard and how it support 4 languages (all part of Spain ofc).

Check it out from this link :
la-leaderboard/la-leaderboard
  • 1 reply
ยท
alielfilali01ย 
posted an update 6 months ago
alielfilali01ย 
posted an update 6 months ago
view post
Post
603
Are the servers down or what ? Am i the only one experiencing this error :
HfHubHTTPError: 500 Server Error: Internal Server Error for url: https://huggingface.co/api/datasets/...../)

Internal Error - We're working hard to fix this as soon as possible!
  • 2 replies
ยท