mmrag-ckpts

mm-RAG checkpoints — indirect-feedback retriever RL

LoRA adapters for GME-Qwen2-VL-2B trained by RL on a frozen VLM reader's answer success, with no relevance labels ([no-gold]: pure top-N policy pools, the only supervision being the gold answer inside the reward). One subfolder per run; the folder name is the full training spec.

Naming: <profile>_<algo>-<reward>_<gold|nogold>_lr_bs_N_cc_steps_rows_seed

Data-scaling ladder

Fixed compute (500 steps x batch 4 => ~2,000 queries visited) with the training pool size varied 12.5k -> 200k. Evaluated on InfoSeek: L1 over 422,378 passages (3,000 queries), L2 top-5 with a Qwen2.5-VL-7B reader on 1,500 queries, v2 token-boundary cover-EM.

Reference points on the identical protocol — GME-2B zero-shot 0.3080, tuned relevance-SFT 0.3100, the released 45k no-gold headline checkpoint 0.3393.

Finding (updated). More training data does not help. From 12.5k to 200k rows at fixed 500 steps: 0.3433 / 0.3487 / 0.3467 / 0.3287 / 0.3480 — a 16x data increase spans 0.0200 while the SFT-to-RL gap is +0.029-0.039. The mechanism is a sampling ceiling: at 500 steps x batch 4 only ~2,000 queries are ever visited, so rows past ~10k are never sampled. This is not a truncated compute curve — tripling the budget at fixed pool makes the model worse (0.3413 -> 0.3193 at 41k; 0.3407 -> 0.3220 at 8k), so the recipe is already at or past its optimum on both axes. The 25k cell replicates at 0.3462 +/- 0.0019 (3 seeds), slightly above the full-data headline config — a shallow interior optimum, not a plateau.

Generalization. Evaluated on 37 MMEB image tasks after InfoSeek-only training. MMEB's OVEN task shares query images with InfoSeek (67/100 sampled), so the 36-task macro excluding it is primary: base 0.4331, relevance-SFT 0.4328 (-0.0003), annotation-free RL 0.4414 (+0.0083). By meta-task, SFT trades VQA-style embedding away (-0.021, worse on 7 of 10 tasks, better on none) for classification (+0.027), while the annotation-free adapter improves exactly the VQA-shaped capability (+0.018, better on 8 of 10) — the capability its answer-success reward optimizes.

Caveat: one seed per ladder cell except 25k (3 seeds).

Full write-up: Irisicy4/VLM2Vec-in @ mmrag