dataset_info:
features:

• name: problem_idx
  dtype: int64
• name: answer
  dtype: string
• name: file_name
  dtype: string
• name: image
  dtype: image
• name: competition
  dtype: string
• name: problem
  dtype: string
• name: problem_type
  list: string
• name: formal_statement
  dtype: string
• name: source
  dtype: string
• name: title
  dtype: string
• name: authors
  dtype: string
• name: points
  dtype: float64
• name: grading_scheme
  dtype: string
• name: sample_solution
  dtype: string
  splits:
• name: train
  num_bytes: 16300548
  num_examples: 560
  download_size: 16102777
  dataset_size: 16300548
  configs:
• config_name: default
  data_files:
  • split: train
    path: data/train-*
    license: cc-by-nc-sa-4.0

Paper Data

This dataset is a merged collection of mathematical problem data used for evaluation. It combines final-answer problems, proof-writing problems, refutation-style problems, image-only multiple-choice problems, and Lean formalization problems into one public dataset repository.

The rows come from multiple competition or benchmark splits. Because the source formats differ, many columns are intentionally nullable. A null value means that the field does not apply to that row or was not available in the source data.

Included Splits

The competition column identifies the source split for each row. The combined upload includes:

• kangaroo/kangaroo_2025_1-2
• kangaroo/kangaroo_2025_3-4
• kangaroo/kangaroo_2025_5-6
• kangaroo/kangaroo_2025_7-8
• kangaroo/kangaroo_2025_9-10
• kangaroo/kangaroo_2025_11-12
• aime/aime_2026
• hmmt/hmmt_feb_2026
• arxivlean/march
• arxiv/december
• arxiv/january
• arxiv/february
• arxiv/march
• arxiv_false/february
• arxiv_false/march
• usamo/usamo_2026
• putnam/putnam_2025
• imc/imc_2025
• miklos/2025
• apex/apex_2025
• apex/shortlist_2025

Schema

Row Types

Final-answer rows contain a text problem and an answer. These include sources such as AIME, HMMT, ArXivMath, Apex, and Kangaroo answer labels. For Kangaroo rows, the problem statement is stored in the image column rather than problem.

Proof-writing rows contain a problem, a grading_scheme, and usually points. They may also include sample_solution. These rows generally have answer = null because they are evaluated by rubric rather than exact answer matching.

Lean rows contain a natural-language problem, a Lean formal_statement, and the same formal statement in answer for compatibility with final-answer style loaders.

Refutation-style rows contain statements that are intended for proof or disproof evaluation. The grading_scheme describes how responses should be judged.

Loading Example

from datasets import load_dataset

dataset = load_dataset("Anon13576245/paper-data", split="train")

row = dataset[0]
print(row["competition"], row["problem_idx"])
print(row["problem"] or row["file_name"])

Notes

• The dataset is intentionally heterogeneous. Consumers should filter by competition or by non-null columns before applying task-specific evaluation logic.
• Null values are expected and should not be treated as missing data errors.
• Image rows can be accessed through the Hugging Face Image feature in the image column.
• The dataset card is identity-neutral. Metadata fields such as authors describe original source material where provided.