Hugging Face's logo
pascalmusabyimana's picture
1 81 197

pascalmusabyimana

pascal-maker
Titus-von-Koeller's profile picture Ameeeee's profile picture Gargaz's profile picture
Ā·
https://pascal-maker.github.io/developedbypascalmusabyimana/

AI & ML interests

computer vision, nlp , machine learning and deeplearning

Recent Activity

reacted to nouamanetazi's post with šŸ¤— about 24 hours ago
After training š’š¦šØš„š‹šŒšŸ‘ on šŸ‘šŸ–šŸ’ š‡šŸšŸŽšŸŽš¬ for nearly a month, I've come to realize something most people overlook: š¢š§šŸš«ššš¬š­š«š®šœš­š®š«šž š¢š¬ š­š”šž š¦ššš¤šž-šØš«-š›š«šžššš¤ šŸšššœš­šØš« š¢š§ š‹š‹šŒ š­š«ššš¢š§š¢š§š . šŸ”„ Everyone talks about model architecture and data quality. And yes, those matter immensely. But here's what nobody tells you: when your training run fails at 2 AM because of mysterious šš‚š‚š‹ šžš«š«šØš«š¬, or when your expensive GPU cluster is running at šŸ”šŸŽ% šžšŸšŸš¢šœš¢šžš§šœš², the problem isn't your model. It's most probably a š¦š¢š¬š®š¬šž šØšŸ š­š”šž š”ššš«šš°ššš«šž. šŸ› ļø Questions that seemed simple but had no clear answers: Why is šŒšØš„ š­š«ššš¢š§š¢š§š  š¬š„šØš°šžš« š­š”ššš§ ššžš§š¬šž š¦šØššžš„š¬? Which šš‚š‚š‹ šŸš„ššš š¬ should we actually set? How often should we checkpoint without killing throughput? That's why we built š“š”šž š’š¦šØš„ š“š«ššš¢š§š¢š§š  šš„ššš²š›šØšØš¤ šŸ“–: a complete guide covering everything from model architecture and data curation to the SmolLM3 training marathon, post-training techniques, and crucially, the š¢š§šŸš«ššš¬š­š«š®šœš­š®š«šž š„ššš²šžš« that most teams get wrong. We validated real vs theoretical bandwidth across the entire stack: š‡ššŒšŸ‘ š”š¢š­š­š¢š§š  šŸ‘ š“š/š¬, šš•š‹š¢š§š¤ šŸ’.šŸŽ š«šžšššœš”š¢š§š  šŸ•šŸ–šŸ” š†š/š¬, šš‚šˆšž š†šžš§šŸ’ ššš­ šŸšŸ’.šŸ š†š/š¬. Then we ran collective operations across šŸšŸšŸ– š†šš”š¬ (16 nodes, 8xH100s each) and measured how performance degrades at scale: all-reduce drops from šŸ’šŸ–šŸŽ š†š/š¬ on a single node to šŸ‘šŸšŸŽ-šŸ‘šŸ“šŸŽ š†š/š¬ across 16 nodes. If you've ever wondered why your training runs are slower than they should be, or you're planning to scale up and want to avoid expensive mistakes, this guide might save you weeks of debugging. š“š”šž š’š¦šØš„ š“š«ššš¢š§š¢š§š  šš„ššš²š›šØšØš¤: https://lnkd.in/e5MKXUHS Shared with ā¤ļø by the HuggingFace team
reacted to Kseniase's post with šŸ”„ 2 days ago
11 Fascinating new Policy Optimization techniques Policy optimization (PO) algorithms are central to training AI models with preference-based feedback. In recent weeks, numerous new PO methods have emerged that build on or replace the popular PPO and GRPO, solving their issues. Here are 11 of them: 1. BAlanced Policy Optimization (BAPO) ā†’ https://huggingface.co/papers/2510.18927 Dynamically adjusting the clipping bounds in PPO-style updates to balance positive and negative gradients and prevent entropy collapse 2. Training-Free GRPO ā†’ https://huggingface.co/papers/2510.08191 Instead of using numeric rewards, it compares rollouts semantically to distill useful knowledge as a token prior, which is then applied during inference to guide the modelā€™s behavior 3. Asymmetric Importance Sampling Policy Optimization (ASPO) ā†’ https://huggingface.co/papers/2510.06062 Fixes imbalanced token weighting in LLM training. It flips the importance sampling ratios for positive tokens to correct over- and under-updates, and adds a soft dual-clipping step to keep gradients stable 4. In-Context Steered Policy Optimization (ICPO) ā†’ https://arxiv.org/abs/2510.26519 Uses a modelā€™s own in-context learning ability to guide training with existing data. It combines Mixed-Policy GRPO with Implicit Expert Forcing to expand exploration and adds Expert Region Reject Sampling and Annealed Expert-Bonus Reward Shaping to ensure stability and balanced expert influence 5. Graph-Enhanced Policy Optimization (GEPO) ā†’ https://arxiv.org/abs/2510.26270 Builds a graph of an agentā€™s experiences to understand how different states connect, guide exploration and assign rewards more effectively 6. Information Gain-based Policy Optimization (IGPO) ā†’ https://huggingface.co/papers/2510.14967 Uses the modelā€™s own belief updates to create dense, informative feedback for smoother multi-turn learning Read further below ā¬‡ļø If you like this, also subscribe to the Turing post: https://www.turingpost.com/subscribe
liked a model 4 days ago
meituan-longcat/LongCat-Flash-Omni
View all activity

Organizations

MLX Community's profile picture Bitsandbytes Community's profile picture