README.md

metadata

license: other
task_categories:
  - visual-question-answering
language:
  - en

CLEVRMath — CLEVRMath_rejected

Rejection-sampled from the CLEVRMath train split. This split holds the rejected items — the answer field holds the official ground truth.

rows	25,396
QA pairs	46,380
shards	167
accepted / rejected (whole family)	1,262,493 / 46,380
accept rate	96.5%
verifier	relaxed

The rejected split is training data, not just diagnostics: answer is the official ground truth, and wrong_vlm records what the model said instead.

How the data was produced

A VLM answers every question at temperature 0 with reasoning enabled. Its answer is compared with the official ground truth by the verifier described below; matches go to CLEVRMath_RS_think and CLEVRMath_RS_nothink, non-matches go to CLEVRMath_rejected with the ground truth preserved.

The question field is byte-identical to the prompt the model actually received, including any appended instruction, so the stored training text and the inference-time input cannot drift apart.

How a match is decided

relaxed — ChartQA relaxed accuracy

The official ChartQA metric. If both the prediction and the reference parse as numbers, the answer is correct when it lands within 5% relative tolerance of the reference — chart values are read off axes, so exact digit agreement is not a fair requirement. Otherwise the two strings must match exactly after lowercasing, trimming, and dropping a trailing period.

Second pass: model-as-judge

String verifiers cannot judge free-form answers at all, and they under-credit answers that are correct but differently phrased. Every item the verifier rejected is therefore re-examined by the same VLM in a text-only pass that sees the question, the reference and the candidate — no image, no re-generation, it only grades text that was already produced.

The judge is instructed to compare against the reference and never to solve the question itself, so an answer is not promoted merely because the model believes it is true. It answers CORRECT, WRONG, or UNSURE; numbers must agree, and UNSURE is reserved for references that are themselves unusable or ambiguous. UNSURE is never counted as correct — only a CORRECT verdict moves a row into the accepted splits. WRONG and UNSURE both remain in the rejected split, so an unresolved case is never silently promoted to a pass.

Calibration: run over items the string verifier had already accepted, the judge agreed on 59/60. Spot-checking its overturns shows it discriminates rather than rubber-stamps — it keeps \frac{L}{n} rejected against a reference of \frac{L}{n-1}, and returns UNSURE when the model argues no finite answer exists against a numeric reference.

Schema

• image — HF Image() — renders directly in the dataset viewer
• qa — list of {question, answer, wrong_vlm, reason}
• clean_meta — per-row provenance and the cleaning policy that admitted it
• image_sha256 — content hash of the image

Examples

Three rows taken straight from this split. Images are the original files as stored in the dataset, and long fields are shown in full.

Example 1

field	value
question	Question: Subtract all big blue things. Subtract all large rubber objects. How many objects are left? Answer in short without any other content.
answer	0
reason	wrong_answer

wrong_vlm (1 chars)

2

Example 2

field	value
question	Question: Add 1 matte objects. How many objects exist? Answer in short without any other content.
answer	10
reason	wrong_answer

wrong_vlm (1 chars)

9

Example 3

field	value
question	Subtract all gray objects. Subtract all gray rubber cylinders. How many objects are left? Answer in short without any other content.
answer	4
reason	wrong_answer

wrong_vlm (1 chars)

5

Cleaning applied before rejection sampling

• Provenance check. Each QA item records where it came from; items whose recorded source does not resolve to the image they are attached to are dropped.
• Task purity. The pool's InfinityMM slice assigns samples to a family by image match rather than by dataset origin, which pulls unrelated general-VQA questions into benchmark families. Families where that content measured off-task keep only their genuine benchmark QA.
• Multiple-choice normalisation. Where options had been flattened into prose, they are re-rendered as a lettered block and the reference is rewritten to the bare letter, so the task is well posed and the answer is unambiguous.