d0rj/mathematics_dataset

Mathematical Reasoning Dataset and RL Environments

A bilingual synthetic school-level mathematics dataset, with an executable single-turn RL/evaluation environment for all 70 modules. The four-column Parquet data remains usable with datasets; the environment package lives in rl/. Adapters support Verifiers v1, Harbor, OpenEnv, NeMo Gym and AgentEnv through a shared checker and bounded task selections. See framework installation and launch commands; these tags do not imply that Hugging Face hosts a running server or that the whole corpus was exported into each framework's task format.

Data

Language split	Rows
en	112,709,888
ru	12,699,944

English data originates from the official DeepMind Mathematics Dataset release. Russian questions were generated independently with a translated generator; the language splits are not paired translations. Exact historical Russian generation flags have not been established for these published files, so this card does not assert a command inferred solely from row counts.

Each row has exactly four string columns:

from datasets import load_dataset

ds = load_dataset(
    "d0rj/mathematics_dataset", "default", split="ru", streaming=True,
    revision="d9f78d0665b3a13dad21c6830b9b358c8a25ae74",
)
print(next(iter(ds)))
# question, answer, category, module

There is no language column in Parquet: select the en or ru split. category separates training difficulty and evaluation:

Category	Intended use
train-easy, train-medium, train-hard	Training / curriculum
interpolate	In-distribution evaluation
extrapolate	Extrapolation evaluation

The eight module families are algebra, arithmetic, calculus, comparison, measurement, numbers, polynomials and probability. See the complete 70-module catalog. Language splits do not replace train/eval selection. Do not mix the two evaluation categories into RL training.

Install the RL environment

Use Linux, macOS or WSL for full execution: the pinned upstream Verifiers runtime requires POSIX fcntl. The standalone checker can also run on Windows.

uvx --from huggingface-hub hf download d0rj/mathematics_dataset \
  --repo-type dataset --include 'rl/**' --local-dir mathematics-env
cd mathematics-env/rl
uv sync --locked --extra verifiers
uv run mathematics-rl catalog

This downloads environment files, not the entire corpus. Package installation is required for this custom v1 taskset; use these commands rather than interpreting the generic library menu as a preinstalled hosted runtime.

Run an evaluation or select training tasks

Point the environment at an OpenAI-compatible model endpoint, with its credential configured locally in LOCAL_LLM_API_KEY when required:

uv run --extra verifiers vf-eval mathematics-rl \
  --env.taskset.language ru --env.taskset.mode interpolate \
  --env.taskset.modules '["arithmetic__add_or_sub"]' \
  --model YOUR_MODEL --client.base-url http://127.0.0.1:8000/v1 \
  --client.api-key-var LOCAL_LLM_API_KEY -n 20 --no-push

Defaults: one agent, one tool-less null turn, local subprocess runtime, and dataset revision d9f78d0665b3a13dad21c6830b9b358c8a25ae74. The loader streams with Parquet filters. Use mode=train for training categories or mode=extrapolate for the separate extrapolation holdout. A Verifiers v1-compatible trainer supplies weight updates; this repository does not start training itself. --no-push keeps evaluation results local instead of uploading them to Prime.

Reward

The evaluator keeps the reference answer private from the solver prompt and scores the final response with reward 1 for correct, 0 otherwise, without an LLM judge. The answer is the last box outside completed thinking, or the last nonempty line. Checks cover exact numbers/fractions, symbolic polynomials, root sets, ordered lists, booleans, time, base numerals, and task-specific factorized/expanded/simplified forms.

For factorization and expansion, equivalence alone is insufficient: copying the original expression does not satisfy the required form. Composed comparisons require preserving the original named-value labels, as specified in the solver prompt. See the full grading contract for accepted syntax and restrictions.

The mathematical parser uses a bounded grammar, not Python execution. A disposable worker enforces a default 3-second deadline per reference check and candidate. Malformed or timed-out source references fail as source errors; model parse errors and timeouts are recorded separately in traces.

Verification and limitations

Version 0.2.1 includes a reproducible local Docker check for all five framework integrations, using pinned SDKs, native clients/servers and actual Harbor agent and separate-verifier containers. See Docker instructions and the recorded container results. The checks cover all 70 modules in both languages with scripted answers; they do not measure model quality.

Run the committed regression suite or bounded reference audit after installation:

uv run pytest tests -q
uv run mathematics-rl self-check --source tests/fixtures/generated_ru.ndjson \
  --language ru --mode all --limit 0 --require-all-modules --output reference-check.json

See validation evidence. Reference acceptance and framework smoke tests do not independently prove every published answer, absence of duplicates or contamination, model quality, or an improvement from RL. No full-corpus mathematical audit or model-training result is claimed. Keep source/filter/seed configuration fixed for resume; see package documentation for bounded shuffling.

Provenance, citation and license

Original generator: google-deepmind/mathematics_dataset. Original paper: Analysing Mathematical Reasoning Abilities of Neural Models, Saxton, Grefenstette, Hill and Kohli, ICLR 2019. Data, translated generator and this environment are Apache-2.0; preserve attribution. Verifiers is a separately installed MIT-licensed dependency. See LICENSE and NOTICE.