davanstrien/qwen35-4b-iconclass-grpo-v5full
The davanstrien/qwen35-4b-iconclass-grpo-v5full is a 4.5 billion parameter Qwen3.5-4B-VL model, developed by davanstrien, specifically fine-tuned for Iconclass classification. This model utilizes a GRPO reward-ablation checkpoint, incorporating a reward configuration of recall, validity, count, and diversity. It aims to improve hierarchical-F1 scores for Iconclass classification, building upon the base davanstrien/qwen35-4b-iconclass-vlm.
Loading preview...
Model Overview
The davanstrien/qwen35-4b-iconclass-grpo-v5full is a 4.5 billion parameter model based on the Qwen3.5-4B-VL architecture, developed by davanstrien. It is a GRPO reward-ablation checkpoint derived from davanstrien/qwen35-4b-iconclass-vlm, specifically designed for Iconclass classification tasks.
Key Characteristics
- Reward Configuration: This model was trained using a reward bundle that includes
recall,validity,count, anddiversitymetrics, in an experiment to see if this richer reward structure outperforms plain hierarchical-F1. - Performance: On a 40-image test set, the model achieved a completeness-corrected H-F1 score of 61.8% for Iconclass classification.
- Experimental Finding: The experiment concluded that this specific reward tuning approach did not yield significant improvement over simpler hierarchical-F1 methods, suggesting the model's performance is primarily capability-bound rather than reward-tuning limited. The README indicates that "anchored fusion" (as seen in
qwen35-4b-iconclass-sft-brillfull) was a more effective approach. - Training: The model was trained using Unsloth and TRL frameworks.
Intended Use
This model is primarily intended for research and experimentation related to Iconclass classification, particularly for evaluating different reward functions in fine-tuning visual language models. It serves as a checkpoint in exploring advanced reward mechanisms for hierarchical classification tasks.