Dataset + pipeline · v1.4.2

Open Reason

An open, verified dataset for coding, science, mathematics, and human reasoning.

Provenance-aware. License-gated. Independently checked. Reproducible.

Open Reason does not use Reddit as a data source. Why

GitHub repository Hugging Face dataset Small CPU model Medium CPU model Large CPU model XL CPU model thanks.dev

GitHub is the lab

Engineering source of truth: pipeline, schemas, taxonomy, source registry, tests, docs, configs, and small samples. Default branch is main.

theworker02/open-reason

Hugging Face is distribution

Parquet shards and the dataset card, versioned from GitHub Releases. CPU causal LMs trained on the corpus: open-reason-small (~1.3M), open-reason-medium (13,867,008), open-reason-large (91,544,064), and open-reason-xl (443,719,680). None of these is a 1B model. Full shards live on Hugging Face, not GitHub.

datasets/theworker02/open-reason

How to load

Install Hugging Face Datasets and pull a named configuration:

from datasets import load_dataset

coding = load_dataset("theworker02/open-reason", "coding")

Other configs work the same way: pass "reasoning", "science", "mathematics", "human", "education", "core", "verified", or "all".

Build locally from the engineering repo:

git clone https://github.com/theworker02/open-reason.git
cd open-reason
pip install -e ".[dev]"
open-reason sources --approve --apply
open-reason build --config all --seed 42 --out data/release

Configurations

coding · reasoning · science · mathematics · human · education · core · verified · all

coding

Sandbox-tested software tasks (Python, SQL, JavaScript when available).

mathematics

Symbolic and integer-checked problems.

science

Independently recomputed numerical and conceptual items.

reasoning

Structured planning and constraint problems.

human

Teaching, synthesis, and decision support.

education

Curriculum graph plus original tasks from auto-approved sources — never scraped lectures.

core

Quality tiers S and A.

verified

Tier S only (quality.verified after a real check).

all

Union by id.

v1.4.0 dataset snapshot

Counts from the release manifest and dataset card. Rebuild with open-reason build --config all --seed 42 --out data/release. Full Parquet shards are on Hugging Face, not GitHub.

Configuration	Examples	Verified	Human-authored
coding	400	386	0
reasoning	580	580	0
science	527	527	0
mathematics	1050	1050	0
human	289	261	28
education	345	111	0
core	3175	2899	28
verified	2899	2899	0
all	3175	2899	28

This is a foundation, not a web dump. Quality over scale. Third-party course sites are registered but not scraped.

Quality, provenance, verification

Most public “reasoning” corpora are web dumps, unverified generations, or evaluation sets reused as training data. Open Reason is built so every row can answer: Where did this come from? May I use it? Was the answer actually checked?

Provenance

Every example includes a provenance object. Source types include human-authored, synthetic (named generator), and reviewed educational sources. Unknown origin requires unknown_reason and cannot be tier S. Fabricating URLs or commits is a pipeline bug.

Verification

quality.verified is set only after a check ran and passed: coding sandbox (unittest / SQLite / Node), sympy or exact integer math, independent numeric science checks, named constraint checkers. If a check was not run, the flag stays false. Votes and “accepted answers” never verify a row.

Policy

License-gated auto-approve enables original tasks inspired by public curricula or docs. It never copies lectures, never sets verbatim=true for NC/SA/unknown licenses, and never enables Reddit or Quora as a source of truth. Contamination fingerprints are reported, not silently deleted. Case study: Why Open Reason does not use Reddit.

Quality tiers

Tier	Meaning
S	A check ran and passed
A	Reviewed / human-authored, not claimed executed
B	Synthetic, structurally valid
C	Raw (unused)

Pipeline

Source registry
License-policy auto-approve
Original task generation
Normalize / validate / Reddit block
Execute or symbolic check
Deduplicate
Contamination report
Statistics + Parquet / JSONL

Citation

If you use the dataset or pipeline, please cite it. See also CITATION.cff.

@misc{openreason2026,
  title        = {Open Reason: An open, verified dataset for coding, science, mathematics, and human reasoning},
  author       = {Open Reason contributors},
  year         = {2026},
  howpublished = {\url{https://github.com/theworker02/open-reason}},
  note         = {Dataset v1.4.0, project v1.4.2}
}