Eurus-2-RL-Data Math and Coding Training Dataset

## Overview Eurus-2-RL-Data is a curated reinforcement-learning training dataset of mathematics and coding problems, each paired with an outcome verifier (LaTeX ground-truth answers for math; executable test cases for code). The dataset was built by the PRIME-RL team to train reasoning models via RL with verifiable rewards. It contains roughly 450K examples (100K–1M range per HF size category), stored in parquet, with both math and coding splits. ## Schema - `prompt` — string — the problem statement / user query - `answer` / `ground_truth` — string — LaTeX answer (math) or reference solution / test harness (code) - `task` — string — task type indicator (e.g., math, code) - `data_source` — string — origin subset (NuminaMath-CoT, APPS, CodeContests, TACO, Codeforces) - `ability` — string — capability category - `reward_model` — struct — verifier configuration / expected answer format - `extra_info` — struct — auxiliary metadata (problem id, difficulty, etc.) ## Sources - HuggingFace: https://huggingface.co/datasets/PRIME-RL/Eurus-2-RL-Data — License: MIT - Upstream math: NuminaMath-CoT (problems from Chinese high school math through IMO-level competitions) - Upstream code: APPS, CodeContests, TACO, Codeforces - Associated papers: arXiv:2502.01456 (PRIME), arXiv:2412.01981 ## Methodology The PRIME-RL team aggregated math problems from NuminaMath-CoT and coding problems from four standard competitive-programming corpora (APPS, CodeContests, TACO, Codeforces). Each example is paired with an automatically checkable verifier: LaTeX final-answer matching for math and unit-test execution for code. The dataset is filtered for quality and formatted for RL-with-verifiable-rewards pipelines (e.g., PRIME, GRPO, RLOO). ## Known gaps & limitations - Heavy overlap with widely-used benchmark suites (e.g., MATH, HumanEval-style coding) is possible since upstream sources (NuminaMath, APPS, CodeContests) are known to contain or be adjacent to common evals — leakage risk if used to train models later evaluated on those benchmarks. - Math problems skew toward competition style; not representative of applied/real-world math. - Code verifiers rely on test cases from upstream sources, which are known to be incomplete in some APPS/TACO problems (false positives possible). - English-dominant; some math problems are translated from Chinese and translation quality varies. - Source does not publish detailed deduplication or contamination audits; buyers should validate empirically against their eval set. ## Intended use & out-of-scope - IS for: RL fine-tuning of reasoning/coding LLMs with verifiable rewards (PRIME, GRPO, RLOO, RLVR), constructing reward models, curriculum design for reasoning training. - NOT for: training models that will be evaluated on MATH / APPS / CodeContests without contamination checks; general SFT (this is RL-formatted with verifiers, not chat SFT); production code-assistant training without additional safety filtering. _Federated dataset: 2 parquet shards, 1.67 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: Eurus-2-RL-Data: Math & Coding RL Training Dataset with Outcome Verifiers High-quality RL training dataset combining math problems (NuminaMath-CoT) and coding problems (APPS, CodeContests, TACO, Codeforces) with outcome verifiers — LaTeX answers for math and test cases for code. ~450K examples, parquet, MIT-licensed.