File size: 1,412 Bytes
2dd39f4 2d70c17 7113af1 2d70c17 7113af1 1bbf72e 7113af1 2d70c17 6dbfd83 2d70c17 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 | ---
license: mit
---
# case-8b-rsft
**Clinical Agents for Seeking Evidence** - agentic harness learning for
longitudinal medical imaging. The model autonomously plans and calls tools inside a
multi-turn harness (model -> `qwen3_xml` tool call -> environment executes ->
observation -> ... -> final answer); it never answers single-shot. Three role
personas: **Anatomist** (organ segmentation / volumetry), **Radiologist**
(cross-visit image comparison), **Consultant** (report-grounded synthesis).
Paper: *From Given to Gathered Evidence: Agentic Learning for Longitudinal Medical Reasoning*.
arxiv.org/abs/2609.39566
SFT-initialized agent (40% of the training split) - already tool-calling capable;
this is the initialization that the RL models continue from.
- **Training**: mixed single-turn SFT over the three agent roles, assistant-only loss
- **Step**: final = 1515（1 epoch）
- **Base model**: Qwen3-VL-8B-Instruct
- **Format**: merged HF checkpoint (single `model.safetensors` + tokenizer/processor)
- **Serving tip**: run with a tool backend (the released mock-tool replay configs,
or the full LangGraph agents) and start vLLM with
`--enable-auto-tool-choice --tool-call-parser qwen3_xml`
(add `--enforce-eager` when the environment has no full CUDA toolkit)
- **Data**: trained on UK Biobank-derived, access-controlled data; no UKB-derived
data is redistributed here. Research use only.
|