Datasets:
language:
- en
source_datasets:
- open-r1/DAPO-Math-17k-Processed
tags:
- mathematics
configs:
- config_name: en
default: true
data_files:
- split: train
path: data/train-00000-of-00001.parquet
dp_removed_DAPO-Math-17k-Processed
Source: open-r1/DAPO-Math-17k-Processed
Pinned source revision: 31dd309567e3da778038cc87d868b6097a3ccf68. Config: en. Split: train. Retained rows: 14068.
All retained rows and original columns are unchanged, including English problems, solutions, reasoning traces, prompt wrappers and source IDs. No translation, text repair, model generation, or added data columns were applied. Stable preparation IDs are recorded separately in provenance.json.
Training/evaluation overlap removal uses normalization v2, distinct 8-gram coverage per training/evaluation pair, and a threshold of 0.7. Coverage uses the evaluation problem as the denominator. Short evaluation problems use exact normalized substring matching. References: AIME 2024, 2025, 2026; AMC 2023; MATH-500. Repeated training problems within or across training datasets are not separately removed.
Summary: {"original_count": 14116, "removed_count": 48, "remaining_count": 14068, "near_miss_count": 18, "removal_counts_by_eval_set": {"aime_2024": 0, "aime_2025": 0, "aime_2026": 1, "amc23": 27, "math_500": 20}}.
Source license: not specified by the source; no new license is asserted.
The original card is retained in SOURCE_README.md. See decontamination_report.json for removals, matching evaluation problems, coverage scores and retained near misses.