# ── Core environment runtime ──────────────────────────────────────────────── pydantic>=2.0 numpy>=1.24 # ── HTTP server (Hugging Face Space, OpenEnv contract) ────────────────────── fastapi>=0.100.0 uvicorn>=0.23.0 python-multipart>=0.0.6 # ── LLM client (used by inference.py for OpenAI-compatible API rollouts) ──── openai>=1.0 # ── Replay rendering / dashboards ─────────────────────────────────────────── matplotlib>=3.7 imageio>=2.28 Pillow>=9.0 # ── Test suite ────────────────────────────────────────────────────────────── pytest>=7.0 pytest-cov>=4.0 # ── Optional extras for evaluating a trained adapter locally ──────────────── # (Training itself is run on Colab/HF Space JupyterLab; see training/grpo_v2_colab.ipynb, # which installs unsloth + trl + datasets inline. These pins here are purely so # `python scripts/eval_trained_model.py` works in a stock Python venv.) # torch>=2.1 # transformers>=4.45 # accelerate>=0.30 # peft>=0.10 # safetensors>=0.4