v3_sftmerged_grounding_grpo_step1600
GRPO LoRA adapter for chest X-ray phrase grounding, on top of the v3 SFT-merged
Lingshu base (lingshu_v3_sft_merged, Qwen2.5-VL-7B). Global step 1600.
- Base model: local merged base
lingshu_v3_sft_merged (see adapter_config.json)
- Method: GRPO (EasyR1/veRL), dual-expert (MDETR/TransVG) tool-calling rollout
- Reward: overall = 0.1format + 0.9IoU
- Hyperparams: lr 5e-6, kl_coef 1e-1, kl_penalty low_var_kl, enforce_eager, LoRA rank 32 / alpha 64, all-linear, vision frozen
- rollout: n=2, temperature 1.0, tp=4, gpu_util 0.6
Notes
- Best val IoU of this run so far: 0.1816 @ step 1400 (val evals every 100 steps oscillate 0.174-0.182).
- Near step 1600: train format ~1.0, accuracy(IoU) ~0.34, any_expert_rate ~0.58 (MDETR-dominated), reward overall ~0.41.
- base_model in adapter_config.json points to the local merged base path; load against that base.
