davanstrien/qwen35-4b-iconclass-grpo-v5ctl
The davanstrien/qwen35-4b-iconclass-grpo-v5ctl model is a 4.5 billion parameter Qwen3.5-4B-VL variant, developed by davanstrien, specifically fine-tuned for Iconclass classification. This model serves as a control in an experiment to evaluate reward tuning strategies, using only a plain hierarchical-F1 reward. It is designed to classify Iconclass categories, achieving 63.7% completeness-corrected H-F1 on a 40-image test set.
Loading preview...
Model Overview
The davanstrien/qwen35-4b-iconclass-grpo-v5ctl is a 4.5 billion parameter model based on the Qwen3.5-4B-VL architecture, developed by davanstrien. It is a reward-ablation checkpoint derived from davanstrien/qwen35-4b-iconclass-vlm, specifically designed for Iconclass classification tasks.
Key Characteristics
- Architecture: Qwen3.5-4B-VL base model.
- Parameter Count: 4.5 billion parameters.
- Context Length: 32768 tokens.
- Training Objective: Trained with Unsloth + TRL, using a
gt_match(plain hierarchical-F1) reward configuration as a control for experimental comparison. - Performance: Achieved 63.7% completeness-corrected H-F1 on a 40-image test set for Iconclass classification.
Experimental Context
This model was part of an experiment to determine if a richer reward bundle could outperform plain hierarchical-F1 for Iconclass classification. The results indicated no significant improvement over the plain gt_match reward, with all tested variants performing within a 61–64% range. This suggests that the model's performance is capability-bound rather than limited by the reward tuning approach. For improved results, the README points to anchored fusion as a more effective strategy, as demonstrated by qwen35-4b-iconclass-sft-brillfull.