open-reason / README.md
theworker02's picture
Update Open Reason v1.4.0 dataset card with real counts
6a6c605 verified
|
Raw
History Blame Contribute Delete
6.72 kB
metadata
pretty_name: Open Reason
license: apache-2.0
task_categories:
  - text-generation
  - question-answering
language:
  - en
tags:
  - reasoning
  - coding
  - mathematics
  - science
  - education
  - provenance
  - evaluation
size_categories:
  - 1K<n<10K
configs:
  - config_name: coding
    data_files:
      - split: train
        path: data/release/coding.parquet
  - config_name: reasoning
    data_files:
      - split: train
        path: data/release/reasoning.parquet
  - config_name: science
    data_files:
      - split: train
        path: data/release/science.parquet
  - config_name: mathematics
    data_files:
      - split: train
        path: data/release/mathematics.parquet
  - config_name: human
    data_files:
      - split: train
        path: data/release/human.parquet
  - config_name: education
    data_files:
      - split: train
        path: data/release/education.parquet
  - config_name: core
    data_files:
      - split: train
        path: data/release/core.parquet
  - config_name: verified
    data_files:
      - split: train
        path: data/release/verified.parquet
  - config_name: all
    data_files:
      - split: train
        path: data/release/all.parquet

Dataset Card for Open Reason

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

Open Reason is a provenance-aware corpus plus a reproducible pipeline. It is intended for training and evaluating systems on coding, mathematics, science, structured decision-making, and human problem solving.

Open Reason does not use Reddit as a data source. Quora is not a primary source of truth. Case study: docs/why-not-reddit.md.

Supported tasks

  • Code generation, debugging, SQL, systems simulations, packaging, and defensive validation
  • Structured reasoning (planning, constraints, causal and temporal problems)
  • Mathematical problem solving with symbolic or integer checks
  • Scientific calculation, modeling, and experimental-design counts
  • Teaching, explanation, and synthesis (human-authored)
  • Curriculum-aligned education tasks with concept ids and education levels

Languages

Prompts and solutions are English. Verified coding languages in v1.4.0: Python, SQL, JavaScript (when the sandbox can run them). Other languages appear as original concept tasks and are not marked verified.

Source information

Kind How to recognize v1.4.0
Human-authored provenance.source_type = human_authored Teaching, synthesis, qualitative items
Synthetic provenance.source_type = synthetic plus generator Math, science, most reasoning/coding, curriculum
Source-derived open_source / community with provenance URL/commit GitHub-permissive original tasks; Stack Overflow seeds (verbatim=false)
Verified quality.verified = true and verification.passed = true Coding sandbox, sympy, numeric, constraint checks
Unverified quality.verified = false Reviewed teaching and misconception items (tier A)

Never treat synthetic rows as human-authored. Never treat unverified rows as executed.

Licensing

Original dataset content and pipeline: Apache 2.0. Per-row provenance.license_spdx is authoritative for upstream GitHub/SO snippets.

Provenance

See docs/provenance.md. Unknown origin requires unknown_reason.

Preprocessing

Unicode NFKC, newline normalization, trimmed lists, task_type slugging. Meaning is not paraphrased.

Deduplication

Exact SHA-256 of canonical fields, normalized prompt/answer hashes, 64-bit simhash. Stats in the release manifest.

Contamination controls

configs/denylist.yaml fingerprints known eval sets. Hits are reported, not silently deleted. --strict fails the build on hits. Hold out benchmarks/ from training.

Quality controls

Schema, SPDX allowlist, Reddit rejection, Quora-as-source rejection, PII heuristics, sandbox/sympy/numeric checks, community-votes-are-not-verification. Tiers S/A/B/C: docs/quality.md. Reddit case study: docs/why-not-reddit.md. evidence_confidence is not a claim of truth.

Intended uses

Research on reasoning and code models; filtering by domain, language, tier, and license; evaluation using the separate benchmarks/ suite.

Limitations

Small v1.4.0 corpus (~3.2K rows); still English-centric; verified coding languages limited to sandbox runtimes; teaching items are not executable oracles unless a numeric/sympy/sandbox check exists; third-party educational sites are registered but not scraped; denylists cannot be complete.

Bias considerations

Synthetic generators encode the authors' choice of topics (software engineering, STEM calculations, operational triage). They under-represent many human domains and languages.

Ethical considerations

No Reddit/social dumps. Case study: docs/why-not-reddit.md. Defensive security only. Minimize PII. Do not present this as a universal "human reasoning" sample.

Maintenance

Issues and PRs: https://github.com/theworker02/open-reason
Hugging Face dataset: https://huggingface.co/datasets/theworker02/open-reason
Small CPU model (1.3M): https://huggingface.co/theworker02/open-reason-small
Medium CPU model (
13.9M): https://huggingface.co/theworker02/open-reason-medium
Releases are immutable; GitHub tags map to Hub revisions. Fixes ship in a new version. Shards are not stored in the GitHub git tree.

Citation

See CITATION.cff and the README BibTeX entry.

v1.4.0 snapshot

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.