Open Trajectory Gym v0.1.0
First public release of Open Trajectory Gym — an open-source platform for post-training LLMs on multi-turn tool-use trajectories.
What's Included
- 3-stage training pipeline: SFT (TRL) → Online RL (SkyRL) → GEPA (DSPy)
- CyBench benchmark: 40 CTF challenges with 8-signal reward function
- 4 model configs: Qwen3.5-27B, Qwen3.5-9B, Qwen3.5-4B, Devstral-24B
- Bring-your-own: Agent, model, benchmark, and reward extension points
- BoxPwnr integration: Reference agent with adapter protocol
- 12 CLI commands: trajgym-train, trajgym-eval, trajgym-agent, trajgym-export, and more
- Docker support: Multi-stage Dockerfile with SFT and Online RL targets
- 20 compatibility patches: SkyRL 0.3.1 + vLLM 0.16 + Ray 2.54
Requirements
- Python 3.11+, Docker, NVIDIA GPU (24GB+ for small models, 2x 140GB+ for Qwen3.5-27B)
- See docs/quickstart.md for setup
Full Changelog: https://github.com/westonbrown/open-trajectory-gym/commits/v0.1.0