Mathematics RL
A mathematics RL environment for Verifiers v1, Harbor, OpenEnv, NeMo Gym and
AgentEnv, covering all 70 modules of
d0rj/mathematics_dataset.
The Verifiers API is provided by pinned verifiers==0.4.0. Each episode is one
tool-less model response followed by a deterministic exact typed reward in {0, 1}.
English and Russian use the same scoring rules and separate language splits.
See FRAMEWORKS.md for the four additional adapters, pinned SDK
versions, bounded exports, native servers/clients and AgentEnv episode loops.
Install each framework in a separate environment: Verifiers' MCP major version
conflicts with the other SDKs. The base package contains only the checker/loader.
Installation
Commands below run from the package directory: cd environments/mathematics_rl
in the source checkout, or cd mathematics-env/rl after the Hub download.
Full Verifiers execution requires Linux, macOS or WSL: the pinned upstream runtime
imports fcntl. The standalone checker and its unit tests also work on Windows.
Scoring needs no GPU, Docker, hosted sandbox or paid service.
uv sync --locked --extra verifiers
uv run mathematics-rl catalog
uv run mathematics-rl grade \
--module arithmetic__div --question 'Divide 1 by 2.' --reference 1/2 --answer '0.5'
For full integration tests from PowerShell (adjust the checkout path):
wsl -d Ubuntu-24.04 -- bash /mnt/d/Projects/mathematics_dataset/scripts/run_rl_tests.sh
The script keeps Linux dependencies in .venv-wsl, separately from Windows .venv.
Evaluation
Use an OpenAI-compatible model endpoint. Configure credentials locally through
the environment variable named by --client.api-key-var. Do not place credentials
in commands, TOML files, dataset cards or traces. Local endpoints may accept EMPTY.
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
The default MathematicsEnv is one agent with the null harness, one model turn
and local subprocess runtime. It provides no tools. Override the runtime with
normal Verifiers configuration when needed. Use --no-push for local results;
upstream otherwise defaults to uploading completed evaluations to Prime Intellect.
--env.taskset.mode train selects only train-easy, train-medium, train-hard.
interpolate and extrapolate are separate evaluation modes. For example,
--env.taskset.categories '["train-hard"]' selects one difficulty within train;
cross-mode categories are rejected. See MODULES.md for all module names.
The pinned dataset revision is d9f78d0665b3a13dad21c6830b9b358c8a25ae74.
Overriding it requires another full commit hash, never a moving branch. The loader
uses lazy streaming with Parquet category/module predicate pushdown. Choose modules
for small pilots; the existing 125M rows do not need to be rewritten.
Bounded sampling and resume
-n bounds consumed tasks. Do not use unbounded -s over the whole corpus: the
framework's exact shuffle materializes its input. Bound it first:
uv run --extra verifiers vf-eval mathematics-rl \
--env.taskset.language en --select.include.idx 0:1000 -s -n 20 \
--model YOUR_MODEL --client.base-url http://127.0.0.1:8000/v1 --no-push
Alternatively use --env.taskset.shuffle-buffer 1000 --env.taskset.seed 42 for
a bounded approximate streaming shuffle. Preserve revision, filters, seed and
buffer size when resuming. IDs combine source/config, source position and content
hash: repeated identical rows have distinct keys, and subsequent CLI selection
preserves IDs. Changing source filters can change IDs; content_hash supports
cross-selection duplicate analysis.
Offline sources and reference checks
--env.taskset.source rows.ndjson (or .parquet) reads the same four string
columns locally. A local file may include language; otherwise it is inferred
from the selected language. On the Hub, language comes from the split name.
Local examples are marked local, not incorrectly attributed to the Hub revision.
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
This checks reference/verifier compatibility, not independent correctness of every reference. Invalid source references fail loudly instead of becoming model errors. Split separation does not prove that duplicate questions or cross-language mathematical overlap are absent.
Reward and RL
GRADING.md specifies extraction, accepted notation, form requirements,
resource limits and the named-value comparison convention. References stay in
immutable TaskData on the evaluator and are not inserted into the solver prompt.
Traces include correct reward, accuracy, parse_error, grading_timeout,
answer kind and grading version.
Use the exported environment with a Verifiers v1-compatible trainer and mode=train.
The trainer supplies trajectories, tokenization, log-probabilities, grouping and
weight updates. This package implements no optimizer and claims no training gain.
Keep both evaluation categories out of the training mixture.
Tests and publication
To check every framework with its pinned SDK in local Linux containers, run
python integrations/docker/run.py --output /absolute/path/to/new-results.
See DOCKER.md for requirements, native protocol coverage and logs.
uv run pytest tests -q
uv run ruff check src tests
uv build .
See VALIDATION.md for evidence and fixture provenance. The repository's
scripts/prepare_hf_release.py builds an allowlisted, hashed release tree with the
corrected card and this package under rl/. Preparation and the publisher's default
dry run do not modify the Hub. See publication/README.md in the source checkout.
Apache-2.0; original generator and source data attributed to DeepMind. See LICENSE and NOTICE.