Why this model exists (research finding)

Built to test whether fixing truncated training labels lifts the iconclass classifier past its ~25% recall ceiling. It does not: training converged well (eval_loss 0.47), but on the clean 788-image full-label test it scores H-F1 45.3 / hier-recall 46.4 / code-recall 25.6 — recall unchanged vs models trained on truncated labels.

Conclusion: the 4B is capability-bound (identifying the right codes), not label-bound — neither reward tuning nor label completeness moves it.

The approach that did improve results is anchored fusion: use this model as a precision anchor, then a graded VLM-judge gates in extra codes from semantic retrieval. On the same clean test that lifts results to H-F1 47.5 / hier-recall 57.6, with zero extra training.

• Base: unsloth/Qwen3.5-4B-Base
• Recommended use: as the anchor in the anchored-fusion pipeline (best recall).

Trained with Unsloth + TRL.