Correction to the previous commit's rationale: it claimed no 1k copy existed and the ladder was therefore incomplete. That was wrong -- the repo did have 1k/MMT-Bench_VAL_MI.tsv; my local mirror had simply not downloaded it. The removal still matches the intent (MMT-Bench_VAL_MI was excluded from the study and is not referenced by eval_results/), and leaving 4 of 5 rungs deleted would be worse than either extreme. Restorable from history.
Pruned to what the resolution/SR study could actually score. None of these appear in this repo's eval_results/, so no recorded run depends on them:
InfoVQA_TEST (5 rungs, 82MB) every answer blank -- unscoreable by anyone
WildVision (5 rungs, 22MB) every answer blank (arena set, no references)
AMBER (5 rungs, 107MB) single constant label 'yes', 7 questions
WildDoc (5 rungs, 5MB) single constant label 'below', 2 questions
MMT-Bench_VAL_MI (4 rungs, 368MB) excluded from the study; no 1k copy so the ladder is
incomplete; MMT-Bench_VAL covers the family and IS
referenced by eval_results runs, so it stays.
KEPT despite being excluded from the study: CoreCognition and MMT-Bench_VAL, because
eval_results/ holds checkpoint-4209_* and SFT-no_crop-bs128-seed43_* runs on both.
Add HRBench8K rotation fix: pool shipped only cycle_category=0 (all answers 'A')
HRBench8K in this pool kept only one CircularEval rotation, so all 151 answers are 'A' and always-answering-A scores 100%. HRBench8K is the only affected benchmark (verified: sole file with a cycle_category column, a constant answer, and index gaps of 4).
HRBench8K_rotation_fix.tsv re-deals one rotation per question, balanced (A:38 B:38 C:38 D:37, majority class 25.2% = chance). Join on `index` and override answer/A/B/C/D; the base64 `image` column is unchanged, so the 2.3GB of images need no re-upload. Question/options/answer are identical across all five rungs (verified 151/151), so one file serves every resolution. Correct answer CONTENT is unchanged for all questions; original_answer and cycle_category preserve provenance.
Upload jsonl_release/README.md with huggingface_hub
