Tag
Explore by tags
Provides ~1.3 trillion tokens of web pages filtered for educational quality using an LLM-trained classifier; includes per-Crawl configs, smaller random samples (10B/100B/350B tokens), and the classifier code and model for reproducible filtering.
A community speedrun to train a 124M GPT as fast as possible on 8 H100s, all chasing a fixed 3.28 FineWeb loss. Successive records cut the run from llm.c's 45 minutes to under 1.4, mostly via the new Muon optimizer rather than more hardware.
Trains a sub-100M-parameter LLM from scratch — pretraining, SFT, LoRA, DPO/RLHF, and distillation, sized from ~26M up to ~100M-plus dense and MoE. Headline figure: the ~64M minimind-3 variant's SFT stage runs 1 epoch in ~2h and ~3 RMB on one NVIDIA 3090.
Trains a 65M-parameter vision-language model from scratch in ~2 hours on one RTX 3090, about 3 RMB (~$0.40) of GPU rental. Connects a frozen SigLIP2 encoder to a small MiniMind LLM via a two-layer MLP projector; full PyTorch code for pretraining and SFT.
Turns PDFs and images into clean Markdown with a 7B vision-language model, keeping tables, equations, handwriting, and multi-column reading order while removing headers and footers. Runs on one 12GB+ GPU at about 1/32 the cost of GPT-4o APIs.
GPU‑accelerated framework for training physically simulated humanoid characters and robots using reinforcement learning and motion imitation. Provides a modular multi‑backend simulator stack, large‑scale multi‑GPU training recipes, built‑in motion retargeting and an ONNX deployment pathway to real robots.
Lets Python developers write tile-based parallel kernels for NVIDIA GPUs, generating CUDA Tile IR while staying close to Python syntax for custom GPU operations.
High-resolution image and video generation codebase and models that run with far lower compute and memory than typical diffusion systems. Uses linear-attention DiT variants, aggressive latent compression, and inference-scaling to support text-to-image (up to 4K), fast one/few-step generation, and efficient video pipelines.
Parses, generates, and filters training data from noisy sources like PDFs and weak QA, then feeds it into LLM pre-training, SFT, RL, or RAG cleaning. Ships 100+ operators and ready-made pipelines for text, reasoning, Text2SQL, and agentic data.
Open-source HybridFlow implementation for RL post-training of LLMs. Decouples control flow from compute so PPO, GRPO, GSPO and DAPO share one dataflow; pairs FSDP/Megatron with vLLM/SGLang rollout and reports 1.5-20x throughput over prior RLHF stacks.
High-accuracy biomolecular structure prediction suite: open-source models (protenix-v2/v1), a benchmark/evaluation toolkit, and a web server for inference. Targets protein/antibody–antigen and ligand-aware predictions with inference-time sampling and constraint support.
A GitHub repository of learning notes and code dedicated to ML + SYS (machine learning systems). It collects tutorials, code walkthroughs and engineering notes on RLHF, distributed training (FSDP, Megatron), inference and scheduling (SGLang, vllm), quantization, CUDA/GPU optimization, system design, and practical engineering.