zbeeb/Staleness-GRPO-DAPO-Math-17k
Staleness GRPO DAPO Math 17k The exact 17,005-row training dataset shared by the staleness-cap-2 Qwen2.5-Math-1.5B, Qwen2.5-3B, and Qwen2.5-Math-7B checkpoints, and the staleness-cap-4 Qwen2.5-Math-1.5B checkpoint. All four training manifests record the same SHA-256 for the training file. Source and processing Derived from the all configuration of open-r1/DAPO-Math-17k-Processed, itself processed from BytedTsinghua-SIA/DAPO-Math-17k. Source revision:… See the full description on the dataset page: https://huggingface.co/datasets/zbeeb/Staleness-GRPO-DAPO-Math-17k.
Staleness GRPO DAPO Math 17k
The exact 17,005-row training dataset shared by the staleness-cap-2 Qwen2.5-Math-1.5B, Qwen2.5-3B, and Qwen2.5-Math-7B checkpoints, and the staleness-cap-4 Qwen2.5-Math-1.5B checkpoint. All four training manifests record the same SHA-256 for the training file.
Source and processing
Derived from the all configuration of open-r1/DAPO-Math-17k-Processed, itself processed from BytedTsinghua-SIA/DAPO-Math-17k. Source revision: 31dd309567e3da778038cc87d868b6097a3ccf68.
Processing started from 17,398 rows, removed 213 canonical duplicate extras, excluded all five conflicting-answer groups (10 source rows; the duplicate-extras count already includes one extra per conflicting group), removed 112 automatically matched evaluation overlaps and 55 manually audited variants, and removed 8 prompts exceeding 1,024 tokens with the training prompt/template. Rows were shuffled with seed 42. The resulting file is copied byte-for-byte from the completed training runs.
Overlap filtering used the retained MATH-500, AMC23, AIME24–26, Minerva Math and text/final-answer OlympiadBench evaluations. These checks do not establish absence of pretraining contamination or exhaustive near-duplicate removal. This repository contains training data; it does not bundle the benchmark test sets.
Load
from datasets import load_dataset
data = load_dataset("zbeeb/Staleness-GRPO-DAPO-Math-17k", split="train")Each record preserves the original training fields, including problem text, accepted final answers, source identifier, benchmark label, and prompt-token count. These are math problems with terminal-answer targets, not generated reasoning traces. Prompt-token counts correspond to the training tokenizer and prompt template, not arbitrary tokenizers.
Integrity
- Rows: 17,005
data/train.jsonlSHA-256:285a7b92a3b5a8efee80bda7506764f7ffd42a195303bc9cb77c399927a2b2c3- Source and filtering provenance: dataset-provenance.json
Attribution and license
Credit to the DAPO authors and the Open R1 processing work. The original DAPO dataset declares Apache-2.0; this redistribution preserves that license and includes its text in LICENSE. The processed upstream card does not add a separate license declaration. See DAPO: An Open-Source LLM Reinforcement Learning System at Scale for the source work.
