Hugging Face's logo

llmc
community
https://github.com/karpathy/llm.c
Activity Feed Request to join this org

AI & ML interests

LLM training in simple, pure C/CUDA
eliebak 
posted an update about 2 months ago
view post
Post
3270
Super excited to announce that our research team at Hugging Face will be doing an AMA on reddit r/LocalLLaMA.

Come ask any questions to the team behind SmolLM, FineWeb and more! And who knows, maybe there’ll be a shiny new release to talk about?

Thursday 4th September, 8AM-11AM PST 🤗

science
eliebak 
posted an update 2 months ago
view post
Post
628
Motif 2.6B tech report is pretty insane, first time i see a model with differential attention and polynorm trained at scale!

> It's trained on 2.5T of token, with a "data mixture schedule" to continuously adjust the mixture over training.
> They use WSD with a "Simple moving average" averaging the last 6 ckpt every 8B token.
> They trained on Finemath, Fineweb2, DCLM, TxT360.
> Lot of details in the finetuning data they used, for instance they used EvolKit and did some "dataset fusion" to have more compressed knowledge into the data.
> They mention they also tried Normalized GPT, QK-Norm and Cross Layer Attention.

Motif-Technologies/Motif-2.6B
1024m 
authored 2 papers 3 months ago
alielfilali01 
posted an update 3 months ago
eliebak 
posted an update 3 months ago
view post
Post
4731
Kimi K2 tech report is full of gems as always. Here are my notes on it:

> MuonClip: Pretty crazy how after 70k the training stabilizes and the QK-clip is basically inactive. There is also no loss in perf with QK-clip which is not trivial at all (at small scale but with aggressive threshold). Also a cool explanation of why muon makes the logit explode in appendix E (tl;dr is that muon makes the singular value of the update matrix higher)
> Sparsity scaling laws to justify their ratio, they have a very solid training infra that allows the model to be trained at this sparsity level, they could have increased even more but as sparsity increases the training becomes less efficient.
> They diminish the number of attention heads to make it more efficient for long context since attention heads are a big bottleneck for long context. They also remove 2 of the 3 "first dense" layers in the dsv3 arch.

With the sparsity and attention heads (divided by 2) they achieve 83% increased flops compared to deepseek v3 arch at 128k.

> Data: Rephrasing is KEY. They do a lot more synthetic data generation and rephrase their corpus to have different styles, for longer documents they do it by chunk. I'm (half) surprised by the fact that ONLY 1 epoch (assuming same number of training tokens I think?) of data rephrased 10 times has better accuracy than 10 epochs of the same data rephrased once.
> They do rewriting for Math and Knowledge, for Math they apply the ShallowMath recipe and instruct the model to rephrase in a "learning note" style
> They talk about diversity and probably have some internal stuff/eval to test that, as always still a bit unclear for me how to properly measure that.

The infra is also very nice, quick summary:
> PP=16 (1F1B schedule, a bit custom), EP=16, zero1
> No FP8 computation but for storage of specific layers, selective recomputation for inexpensive block, activation offloading to CPU
eliebak 
authored a paper 5 months ago
alielfilali01 
authored a paper 5 months ago
1024m 
authored a paper 5 months ago
alielfilali01 
posted an update 6 months ago
1024m 
authored 2 papers 6 months ago
1024m 
authored a paper 7 months ago
eliebak 
authored a paper 7 months ago
mmhamdy 
posted an update 7 months ago
view post
Post
1690
What inspired the Transformer architecture in the "Attention Is All You Need" paper? And how were various ideas combined to create this groundbreaking model?

In this lengthy article, I explore the story and the origins of some of the ideas introduced in the paper. We'll explore everything from the fundamental attention mechanism that lies at its heart to the surprisingly simple explanation for its name, Transformer.

💡 Examples of ideas explored in the article:

✅ What was the inspiration for the attention mechanism?
✅ How did we go from attention to self-attention?
✅ Did the team have any other names in mind for the model?

and more...

I aim to tell the story of Transformers as I would have wanted to read it, and hopefully, one that appeals to others interested in the details of this fascinating idea. This narrative draws from video interviews, lectures, articles, tweets/Xs, and some digging into the literature. I have done my best to be accurate, but errors are possible. If you find inaccuracies or have any additions, please do reach out, and I will gladly make the necessary updates.

Read the article: https://huggingface.co/blog/mmhamdy/pandemonium-the-transformers-story
eliebak 
posted an update 8 months ago
view post
Post
2084
Google just dropped an exciting technical report for the brand-new Gemma3 model! 🚀 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
·
mmhamdy 
posted an update 8 months ago
view post
Post
2768
🎉 We're excited to introduce MemoryCode, a novel synthetic dataset designed to rigorously evaluate LLMs' ability to track and execute coding instructions across multiple sessions. MemoryCode simulates realistic workplace scenarios where a mentee (the LLM) receives coding instructions from a mentor amidst a stream of both relevant and irrelevant information.

💡 But what makes MemoryCode unique?! The combination of the following:

✅ Multi-Session Dialogue Histories: MemoryCode consists of chronological sequences of dialogues between a mentor and a mentee, mirroring real-world interactions between coworkers.

✅ Interspersed Irrelevant Information: Critical instructions are deliberately interspersed with unrelated content, replicating the information overload common in office environments.

✅ Instruction Updates: Coding rules and conventions can be updated multiple times throughout the dialogue history, requiring LLMs to track and apply the most recent information.

✅ Prospective Memory: Unlike previous datasets that cue information retrieval, MemoryCode requires LLMs to spontaneously recall and apply relevant instructions without explicit prompts.

✅ Practical Task Execution: LLMs are evaluated on their ability to use the retrieved information to perform practical coding tasks, bridging the gap between information recall and real-world application.

📌 Our Findings

1️⃣ While even small models can handle isolated coding instructions, the performance of top-tier models like GPT-4o dramatically deteriorates when instructions are spread across multiple sessions.

2️⃣ This performance drop isn't simply due to the length of the context. Our analysis indicates that LLMs struggle to reason compositionally over sequences of instructions and updates. They have difficulty keeping track of which instructions are current and how to apply them.

🔗 Paper: From Tools to Teammates: Evaluating LLMs in Multi-Session Coding Interactions (2502.13791)
📦 Code: https://github.com/for-ai/MemoryCode
mmhamdy 
authored 2 papers 8 months ago
alielfilali01 
posted an update 8 months ago
view post
Post
1095
🚨 Arabic LLM Evaluation 🚨

Few models join the ranking of https://huggingface.co/spaces/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 https://huggingface.co/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.