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

VISIONConcurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The davanstrien/qwen35-4b-iconclass-grpo-v5hyb is a 4.5 billion parameter Qwen3.5-4B-VL model, derived from davanstrien/qwen35-4b-iconclass-vlm. This model was developed by davanstrien as an experimental checkpoint for iconclass classification, specifically testing a GRPO reward-ablation configuration. It focuses on evaluating whether a richer reward bundle (gt_match + count + diversity) improves performance over plain hierarchical-F1 for this specialized visual classification task.

Loading preview...

Model Overview

The davanstrien/qwen35-4b-iconclass-grpo-v5hyb is a 4.5 billion parameter model based on the Qwen3.5-4B-VL architecture. It is an experimental checkpoint derived from davanstrien/qwen35-4b-iconclass-vlm, developed by davanstrien.

Experimental Focus

This model was created as part of an experiment to test the effectiveness of a richer reward bundle in Grouped Reward Policy Optimization (GRPO) for iconclass classification. The specific reward configuration used was gt_match + count + diversity, aiming to improve upon plain hierarchical-F1 (gt_match) performance.

Performance and Findings

  • Completeness-corrected H-F1 (40-image test): 61.7%
  • Verdict: The experiment concluded that this reward configuration showed no significant improvement over plain gt_match for iconclass classification. All tested variants achieved similar scores (61–64%), indicating that the model's capabilities, rather than the reward tuning, were the limiting factor. The README suggests that "anchored fusion" (as seen in qwen35-4b-iconclass-sft-brillfull) was a more effective approach for this task.

Training Details

The model was trained using Unsloth and TRL, building upon the davanstrien/qwen35-4b-iconclass-vlm base.