Image Captioning on DenseFusion (50k Sampled Images)
64.4CapArena Win Rate (vs Base Model)RubiCap-7B-DenseFusion
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| RubiCap-7B-DenseFusionFine-tuning=GRPO, Data=50k DenseFusion2026.03 | 64.4 | 53.2 | 62.21 | 74.51 | 28.58 | 21.15 | 12.8 | 45.26 | |
| Direct-LikertFine-tuning=GRPO, Data=50k DenseFusion2026.03 | 59.2 | 48.2 | 62.01 | 74.9 | 28.38 | 21.24 | 12.75 | 43.81 | |
| NLP Metric-based RLFine-tuning=GRPO, Data=50k DenseFusion2026.03 | 58.4 | 45 | 62.35 | 77.52 | 28.9 | 21.5 | 13.3 | 43.85 | |
| Reference-LikertFine-tuning=GRPO, Data=50k DenseFusion2026.03 | 50.6 | 38.4 | 62.11 | 73.43 | 26.39 | 23.17 | 9.89 | 40.57 | |
| Base Model (Qwen2.5-VL-7B-Instruct)Protocol=Pre-trained, Backbone=Qwen2.5-VL-7B-Instruct2026.03 | 50 | 44.8 | 60.64 | 69.58 | 27.25 | 19.42 | 11.54 | 40.46 |