Sft GitHub Repositories (2026) · Open Source Tools

Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.

Use PEFT or Full-parameter to CPT/SFT/DPO/GRPO 600+ LLMs (Qwen3.8, DeepSeek-V4, GLM-5.1, InternLM3, Llama4, ...) and 300+ MLLMs (Qwen3-VL, Qwen3-Omni, InternVL3.5, Ovis2.5, GLM5.3, Gemma4, Llava, Phi4, ...) (AAAI 2025).

A simple, performant, and scalable Jax LLM!

The complete post-training stack for LLMs, plus RL for any problem. SFT, DPO and multi-turn agentic RL, with evolutionary hyperparameter optimization built in.

A native Mac App for LLM fine-tuning on Apple Silicon — fully on-device, fully open source.

Human-feedback and data-governance infrastructure for RSI research: annotation, review, provenance and frozen dataset candidates.

Train and serve LLMs at extreme speed and massive throughput.