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. Case study: docs/why-not-reddit.md · project site.

Official links

What	URL
GitHub (pipeline, tests, samples)	https://github.com/theworker02/open-reason
Dataset (full shards)	https://huggingface.co/datasets/theworker02/open-reason
Small model (~1.3M params, CPU)	https://huggingface.co/theworker02/open-reason-small
Medium model (13,867,008 params, CPU)	https://huggingface.co/theworker02/open-reason-medium
Large model (91,544,064 params, CPU)	https://huggingface.co/theworker02/open-reason-large
XL model (443,719,680 params, CPU)	https://huggingface.co/theworker02/open-reason-xl
Site	https://theworker02.github.io/open-reason/

Full Parquet/JSONL shards are on Hugging Face, not GitHub. Project license is Apache-2.0 only.

from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

ds = load_dataset("theworker02/open-reason", "all")
tok = AutoTokenizer.from_pretrained("theworker02/open-reason-xl")
model = AutoModelForCausalLM.from_pretrained("theworker02/open-reason-xl")
# or: theworker02/open-reason-large / theworker02/open-reason-medium / theworker02/open-reason-small

Changelog

See CHANGELOG.md for every release.

Version	Date	Notes
v1.4.2	2026-08-18	XL CPU causal LM (443,719,680); dataset unchanged
v1.4.1	2026-08-18	Large CPU causal LM (~91.5M); dataset unchanged
v1.4.0	2026-08-18	Larger verified corpus; small + medium CPU models
v1.3.8	2026-08-18	Apache-2.0 only; train-sized corpus; small CPU causal LM
v1.0.0	2026-08-18	Catalogs in every section, broader original tasks, local 1.0 dataset
v0.4.0	2026-08-18	Policy engine, coverage generation, training pipeline
v0.3.0	2026-08-18	Curriculum auto-approve, original source-tagged tasks, README demos
v0.2.0	2026-08-18	Source registry, knowledge graph, education / core / verified splits
v0.1.0	2026-08-18	Schema, sandbox verification, first seed release

Why it exists

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?

It is a dataset and the pipeline that produces it.

GitHub vs Hugging Face

• GitHub theworker02/open-reason is the lab: pipeline, schemas, taxonomy, registry, tests, docs, configs, and small samples. Default branch is main. Full shards are not in git.
• Hugging Face dataset theworker02/open-reason is distribution: Parquet shards and the dataset card.
• Small CPU model theworker02/open-reason-small is a ~1.3M-parameter GPT-2-style causal LM.
• Medium CPU model theworker02/open-reason-medium is a 13,867,008-parameter GPT-2-style causal LM (CPU, not 1B).
• Large CPU model theworker02/open-reason-large is a 91,544,064-parameter GPT-2-style causal LM (CPU, not 1B).
• XL CPU model theworker02/open-reason-xl is a 443,719,680-parameter GPT-2-style causal LM (CPU, not 1B). Weights are on the Hub, not in git.
• open-reason build writes local data/release/. Those *.parquet / *.jsonl shards are gitignored. A committed sample lives in data/sample/.

See docs/huggingface.md and docs/releases.md.

Quick start

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

Load:

from datasets import load_dataset

coding = load_dataset("theworker02/open-reason", "coding")
math = load_dataset("theworker02/open-reason", "mathematics")
education = load_dataset("theworker02/open-reason", "education")
core = load_dataset("theworker02/open-reason", "core")

How to use it

1. Auto-approve sources

Auto-approve is a license policy, not a scrape. It enables original Open Reason tasks inspired by a source’s public curriculum or docs. It never copies lectures, never sets verbatim=true for NC/SA/unknown licenses, and never enables Reddit.

open-reason sources --approve          # dry run
open-reason sources --approve --apply  # write sources/registry.yaml

2. Generate original examples

open-reason generate --domain education
open-reason ingest --source khan-academy   # original tasks, not copied lessons
open-reason ingest --source reddit         # rejected

3. Load a configuration

from datasets import load_dataset
ds = load_dataset("theworker02/open-reason", "coding", split="train", streaming=True)
for row in ds.take(3):
    print(row["id"], row["task_type"])

Dataset configurations

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

Config	Contents
coding	Sandbox-tested software tasks (Python, SQL, JavaScript when available)
mathematics	Symbolic / integer-checked problems
science	Independently recomputed numerical and conceptual items
reasoning	Structured planning and constraint problems
human	Teaching, synthesis, decision support
education	Curriculum graph + original tasks from auto-approved sources
core	Quality tiers S and A
verified	Tier S only (quality.verified after a real check)
all	Union by id

Auto-approve policy

permissive SPDX + commercial + no share-alike  →  original tasks (verbatim still off until a reviewed crawler)
education / docs with unclear or SA/NC terms   →  original tasks only
Reddit / Quora / prohibited                    →  never
Stack Overflow                                 →  original rewritten seeds only (not verbatim CC BY-SA dumps)

quality.verified is never set from votes, views, or “accepted answer.”

Pipeline

source registry
    → license-policy auto-approve
    → original task generation
    → normalize / validate / Reddit block
    → execute or symbolic check
    → deduplicate
    → contamination report
    → statistics + Parquet / JSONL

open-reason sources
open-reason generate --domain programming
open-reason validate data/release --strict
open-reason statistics --config all
open-reason benchmark

Schema

Every example carries knowledge, task, evidence, solution, verification, provenance, educational position, and quality — not only prompt + answer.

{
  "id": "or-mathematics-synthetic-…",
  "domain": "mathematics",
  "task_type": "algebra",
  "prompt": "…",
  "observations": [],
  "constraints": [],
  "plan": [],
  "solution": "…",
  "answer": "…",
  "verification": {"method": "sympy", "passed": true},
  "provenance": {"source_type": "synthetic", "license_spdx": "Apache-2.0"},
  "quality": {"tier": "S", "verified": true, "evidence_confidence": 0.81},
  "education_level": "high_school",
  "concept_id": "math.algebra"
}

JSON Schema: schemas/.

Quality tiers

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

Licensing

• Project: Apache 2.0
• Per-row provenance.license_spdx records upstream SPDX (GitHub MIT/BSD/Apache snippets, SO-inspired original rows)
• Share-alike and non-commercial third-party text is not relicensed into this release

Statistics

Pipeline version 1.4.0.

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

Rebuild with open-reason build --config all --seed 42 --out data/release. Full tables: data/release/statistics.md.

Limitations

• Quality over scale: this is a foundation, not a web dump
• Auto-approve does not download Khan Academy, MIT OCW, MDN, or Stack Overflow
• Verified coding languages are those the sandbox can run
• Teaching items are not executable oracles
• Benchmark denylists cannot be complete

Documentation

• Project site
• Architecture
• Data sources · Why not Reddit
• Knowledge graph
• Provenance · Licensing · Quality
• Validation · Verification · Sandbox
• Contamination · Releases
• Hugging Face · Dataset card
• Evaluation · Training
• Contributing · Security

Citation

@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}
}

Also see CITATION.cff.

Open. Licensed. Provenanced. Diverse. Verified. Reproducible. Completely free of Reddit.