GRPO reward-ablation checkpoint (Qwen3.5-4B-VL, from davanstrien/qwen35-4b-iconclass-vlm). Part of an experiment testing whether a richer reward bundle beats plain hierarchical-F1 (gt_match) for iconclass classification.

Reward config: gt_match + count + diversity
Result (completeness-corrected H-F1, 40-image test): 61.7%
Verdict: no improvement over plain gt_match (all variants 61–64%, within n=40 noise). Reward tuning is not the lever — the model is capability-bound. The approach that worked is anchored fusion (see qwen35-4b-iconclass-sft-brillfull).

Base: davanstrien/qwen35-4b-iconclass-vlm. Trained with Unsloth + TRL.

davanstrien/qwen35-4b-iconclass-grpo-v5hyb

qwen35-4b-iconclass-grpo-v5hyb

davanstrien/
qwen35-4b-iconclass-grpo-v5hyb