README_PATCH.md

HRBench8K in this pool contains only ONE option rotation — read before using it

<res>/HRBench8K.tsv was sampled from HRBench's CircularEval set keeping only cycle_category == 0, so all 151 answers are A. A model that always answers "A" scores 100%, and any accuracy computed on it as-shipped is meaningless. Signature: the file carries a cycle_category column that is 0 everywhere, and its index values step by 4 (0, 12, 16, 24 …) because every 4th row of the original was taken.

This affects only HRBench8K. Every other benchmark here was checked and is fine (AMBER/WildDoc have a single label because they genuinely are single-label sets, not because of rotation sampling).

Fix: HRBench8K_rotation_fix.tsv

A 50 KB sidecar that re-deals one rotation per question so the labels are balanced (A:38 B:38 C:38 D:37, majority class 25.2% = chance). Join on index and override answer, A, B, C, D:

import csv, sys
csv.field_size_limit(sys.maxsize)
fix = {r["index"]: r for r in csv.DictReader(open("HRBench8K_rotation_fix.tsv"), delimiter="\t")}
for row in csv.DictReader(open("1k/HRBench8K.tsv"), delimiter="\t"):   # any <res>
    f = fix[row["index"]]
    row.update({k: f[k] for k in ("answer", "A", "B", "C", "D", "cycle_category")})
    # row["image"] is untouched

The correct answer content is unchanged for all 151 questions — only which letter carries it moves — and the same question gets the same rotation at every resolution, so cross-resolution comparisons stay valid. original_answer preserves the old value; cycle_category records which rotation the row now is.

Options and answers are identical across all five rungs (verified 151/151 for 256/512/4k/8k against 1k), so this one file serves every resolution.

If you want full CircularEval instead

Generate all four rotations per question (588 rows) rather than one: rotation k places the option at original position i at position (i+k)%4, and the answer letter follows the correct content. That gives mean-per-cycle and strict-4/4 metrics, at 4x the rows.