indirect_caption — side-by-side compare reports

Trained vs base — random + top-by-output-tokens differing rows per benchmark, with embedded images.

• v4_7b_lpan_mixfv vs qwen2_5_vl_7b_base
• v4_7b_lpan vs qwen2_5_vl_7b_base
• v4_3b_lpan vs qwen2_5_vl_3b_base
• v2_1ep_lr5e-6 (PureLlava) vs qwen2_5_vl_7b_base
• v3_last (PureLlava) vs qwen2_5_vl_7b_base

Source CSV: projects/indirect-caption/data/eval_results.csv on exp-record.