10 open-source repos for training small models to do your agent's job
victor_explore · x · 2026-10-17
A curated list of 10 GitHub repos (stars checked Oct 2026) for training small models to handle agent workloads:
- Unsloth (77K★, Apache 2.0): fine-tune or RL-train on one GPU or free Colab, run on a laptop.
- LlamaFactory (75K★): no-code fine-tuning of hundreds of models via a private web UI.
- verl (24K★): heavy-duty RL trainer many agent trainers build on; needs multiple GPUs.
- TRL (19K★): Hugging Face's base library for fine-tuning and RL.
- Agent Lightning (19K★, MIT): Microsoft's tool to RL-train the model inside your existing agent with no rewrite.
- Axolotl (13K★): train from a single YAML file.
- ART (11K★, OpenPipe): RL-train multi-step agents.
- Oumi (9K★): end-to-end data, training, and evals stack.
- rLLM (6K★): agentic RL on any harness, backend, and benchmark.
- NVIDIA DataDesigner (2K★): generate synthetic training data from scratch.
More from coding & agent
- Developer wires Grok bot into Tailscale to manage Plex server from phone — johnlindquist · 2026-10-17
- Poll: 70% say chat is not the final form of agent UI — jonathan_wilke · 2026-10-17
- DeepSeek Harness v0.2.1 nears release with major plugin system overhaul — teortaxesTex · 2026-10-17
- 'Yes, I'm a Full Stack Engineer': The Joke About Juggling Claude, Codex and OpenCode Models — rachittshah · 2026-10-17
- Reddit dev shares checkpoint-and-clear loop for running AI agents for days without losing state — Motiv31 · 2026-10-17
- Building a notes chatbot: why semantic search fails on dates and membership queries — bzarnal · 2026-10-17