birgermoell/oellm-math-rlvr

OpenEuroLLM Math RLVR

One million deterministic, verifier-ready mathematical problems for reinforcement learning with verifiable rewards. The release contains a 760,000-row English depth pool and 10,000 aligned semantic problems rendered in all 24 official EU languages (240,000 rows).

This is a prompt-and-answer rollout corpus, not a chain-of-thought corpus. Model inputs contain only the problem and output-format instruction. Reference answers and verifier contracts remain separate fields.

Design

• —28 procedural generator families spanning arithmetic, algebra, number theory, probability, combinatorics, geometry, statistics, calculus and applied mathematics.
• —Exact integer or reduced-rational answers only in v0.1.
• —Deterministic regeneration from generator_family, generation_seed and difficulty.
• —Stable semantic group IDs keep every aligned language variant in the same split.
• —Duplicate IDs and duplicate language/prompt pairs are rejected.
• —Generator-family contamination groups are explicit so downstream users can create stricter holdouts.
• —Apache-2.0 throughout: no scraped text, benchmark rows or teacher-generated reasoning traces.

The design is inspired by DeepSeekMath's distinction between large-scale math pretraining data and the smaller question bank used for RL. This release focuses specifically on scalable rollout questions and deterministic outcome verification.

Loading

from datasets import load_dataset

ds = load_dataset("birgermoell/oellm-math-rlvr")
print(ds["train"][0])

The messages and ground_truth columns can be passed directly to the OpenEuroLLM RLVR control plane. Never include ground_truth, canonical_answer, parameters_json or verifier metadata in the model prompt.

Columns

Splits

Approximately 95% train, 2.5% validation and 2.5% test, assigned by stable hashing of semantic_group_id. All language variants of a mathematical problem stay together.

For claims about generalization, create a stricter private evaluation set using generator families and problem templates absent from this repository. The included test split measures parameter and language generalization, not unseen-family generalization.

Verification

v0.1 uses two contracts:

• —integer_exact: signed integer equality;
• —rational_exact: equivalence of reduced rational numbers.

Consumers should parse the final \\boxed{...} answer and compare using bounded rational arithmetic. Do not use Python eval.

Multilingual scope

The aligned subset localizes short task instructions and the answer contract while keeping mathematical statements symbolic. It deliberately does not translate synthetic chain-of-thought traces. More natural language-heavy, native-authored problems should be added as a separately audited future source.

Limitations

• —Procedural problems do not substitute for difficult human-authored competition mathematics.
• —The five difficulty levels reflect generator settings, not measured model success probabilities.
• —This version contains no symbolic-expression, interval, unit-aware or proof verifier.
• —Template regularity can be exploited; use family-balanced sampling and add independently sourced tasks.
• —Public test answers are visible and must not be used as the sole release benchmark.

Reproduction and integrity

The repository includes the complete generator and validator under src/. manifest.json records row counts and SHA-256 checksums for every Parquet shard.

python -m pip install -e .
oellm-math-build --output release --rows 1000000
oellm-math-validate release --full

Citation

@dataset{moell2026oellmmathrlvr,
  title = {OpenEuroLLM Math RLVR},
  author = {Moell, Birger and OpenEuroLLM contributors},
  year = {2026},
  url = {https://huggingface.co/datasets/birgermoell/oellm-math-rlvr}
}