| --- |
| license: mit |
| language: |
| - en |
| base_model: |
| - Qwen/Qwen3-VL-8B-Instruct |
| --- |
| # case-8b-rl-abl-rlaif (step 1912, 1 epoch) |
|
|
| **Ablation**: identical to `case-8b-rl` but **without the rubric-judge reward** |
| (the LLM-judge scoring term is removed; the format/outcome terms remain). |
| Paper: *From Given to Gathered Evidence: Agentic Learning for Longitudinal Medical Reasoning*. |
|
|
| - **Training**: GRPO + OPSD, rubric-judge reward disabled |
| - **Step**: 1912 |
| - **Provenance**: `case/rl_ablate_no_judge/hf_step1912` (merged from FSDP shards) |
| - **Base model**: Qwen3-VL-8B-Instruct |
| - **Task**: longitudinal medical-imaging QA; the policy answers by calling tools |
| (view image, compare visits, organ segmentation, report retrieval, skill memory) |
| - **Format**: merged HF checkpoint (single `model.safetensors` + tokenizer/processor) |
| - **Usage**: `vllm serve <this_dir> --trust-remote-code` (add `--enforce-eager` when |
| serving in an environment without a full CUDA toolkit) |
| - **Data**: trained on UK Biobank-derived, access-controlled data; no UKB-derived |
| data is redistributed here. Research use only. |
