# case-env 

**Clinical Agents for Seeking Evidence** - the conda environment package for
`case-8b` training and benchmark evaluation in the paper *From Given to Gathered
Evidence: Agentic Learning for Longitudinal Medical Reasoning*. A single
compressed tar that extracts directly into a ready environment, bit-identical
to the stack that produced all published `case-8b` runs - no dependency
resolution, no rebuild.

- **Contents**: Python 3.10, PyTorch 2.11.0 (CUDA 13.0), vLLM 0.26.0,
  transformers 5.14.1, Ray 2.55.1, LangGraph 1.2.9 (multi-turn tool-agent
  orchestration), verl 0.7.0 (rebuilt for agentic RL training: agent-loop
  state machine and related adaptations, based on upstream verl 0.7.0),
  flash-attn / FLA / FlashInfer stack
- **Size**: 12.6 GB tar; about 12 GB on disk after extraction
- **Install**: the tar's top-level directory is `case/`, and bin/ shebangs are
  pre-rewritten to `/opt/conda/envs/case`, so
  `tar -xf case_env_0904.tar -C /opt/conda/envs` yields a ready env named `case`.
- **Checksum**: sha256 `abf5016f171887449c41fb6d7aaec95e2eaa79439e35fd2443736f3713a8fea8`
- **Verify**: `sha256sum -c case_env_0904.tar.sha256` then
  `/opt/conda/envs/case/bin/python -c "import torch, vllm, verl"`
- **Runtime exports** (already baked into the released run scripts):
  `PYTHONPATH=<repo>/scripts/vllm_patch` and
  `VLLM_USE_FLASHINFER_SAMPLER=0 FLA_CACHE_RESULTS=0 OMP_NUM_THREADS=16`
- **Requirements**: Linux x86_64, NVIDIA driver with CUDA 13.0 support
  (>= 580 series), system gcc (for Triton runtime JIT); ~25 GB free disk during
  install (12.6 GB tar + about 12 GB extracted)
- **Docker alternative**: `case_0904_image.tar.gz` -> `docker load` ->
  `docker run --gpus all -it case:0904 bash`
- **Data**: UK Biobank-derived data is access-controlled and not included in
  this environment. Research use only.