---
license: apache-2.0
language:
- en
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: source
    dtype: string
  - name: problem
    dtype: string
  - name: generation
    dtype: string
  splits:
  - name: train
    num_bytes: 249426162
    num_examples: 15172
  download_size: 109919239
  dataset_size: 249426162
---

# dp_removed_OpenR1-Math-220k

Derived from [open-r1/OpenR1-Math-220k](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k), default/train at `e4e141ec9dea9f8326f4d347be56105859b2bd68`. Original source card: [source_card.md](source_card.md).

15172 problems, 27739 retained traces. One row per problem; `generations` contains ALL retained traces. Parallel arrays refer to those traces; `generation_indices` records original positions. Text is unchanged. Only math-verify-correct traces are included. Explicit length stops and incomplete reasoning are excluded. Missing finish reason policy: `drop`; fallback does not prove the final answer was untruncated and is identified in `completion_evidence`.

Length <= 20480 using `Qwen/Qwen2.5-3B` at `3aab1f1954e9cc14eb9509a215f9e5ca08227a9b`: tokens(problem) + tokens(generation), without special tokens, chat formatting, padding or truncation. Training templates add overhead.

Train/eval overlap removal: normalization v2, distinct 8-grams, coverage >= 0.7 per evaluation problem. Short eval problems use normalized exact substring matching. References: AIME 2024/2025/2026, AMC23, MATH-500; exact pins in provenance.json. No within/between-training deduplication or blanket source-label exclusions. See filter_report.json and decontamination_report.json for counts, matches and near misses.