Code that actually shipped.
Real, production commercial repositories, not synthetic samples or scraped snippets. Every repository ships with contributor history, commit depth, pull request activity and source-to-test coverage.
Available as- 01 · Dataset
- 3B+ lines with commit, PR and review history
- 02 · Environment
- Agentic engineering tasks, engineer-graded
Licensed, stocked, ready to ship.
Plus custom collection for what doesn't exist yet.
Across licensed private production repositories
Real commercial systems, not forks or tutorials
Review and iteration signal, not just commits
From Java and TypeScript to Dart, Go and Kotlin
Commits on a single repository
Continuous commit history on one repository
Metadata, not just source files.
Every repository carries the same fields, so the collection is filterable by language, depth, activity and test coverage.
- Repository identity & activity
- A unique project code per repository, with years active tracked from first to most recent commit.
- Contributor footprint
- Contributor counts per repository, capturing team size and collaboration density over the project's life.
- Language composition
- Primary language plus a full percentage breakdown across every language and framework used.
- Pull & merge request history
- Total and merged PR/MR counts per repository: real review and iteration signal.
- Commit history depth
- Full commit counts per repository, with some histories exceeding 40,000 commits.
- Source vs. test coverage
- Source and test file counts tracked separately, useful for test generation and coverage-aware training.
Repository-level manifest. Field names are illustrative.
From owner licence to review.
One file from repository lic_0193, through the five stages every repository goes through before it ships.
License
Repositories come from the companies that own them, under a licence that permits training. Forks, tutorials and toy projects are excluded.
- Who
- Repository owner
- Gate
- No licence, no ingest
Scrub
Secrets, credentials, personal emails and internal hostnames are found and replaced across every commit in history, not just the latest.
- Who
- Scanner + engineer
- Gate
- Zero live secrets in any commit
Structure
Commits, pull requests, review threads and tests are joined, so each change carries its discussion and the tests that cover it.
- Who
- Data engineering
- Gate
- Every PR keeps its review thread
Engineer review
Engineers check that changes build and tests pass, and label what the task needs: accepted or rejected, and why.
- Who
- Senior engineers
- Gate
- Builds and tests pass, or flagged
Package
Shipped as git bundles or flattened records, with a manifest per repository and a datasheet for the delivery.
- Who
- Delivery
- Gate
- Manifest + datasheet with every drop
Built for frontier model tasks.
Scope, review criteria and delivery format are agreed per engagement, from the first sample to the final eval.
- Code generation
- Pre-training and SFT on production code rather than tutorials.
- Agentic software engineering
- Real bug-fix sequences and review threads for coding agents.
- Code review & judgment
- Accepted and rejected changes with the reasons attached.
- Test generation
- Source paired with its tests for coverage-aware training.
- Migration & refactoring
- Long histories that show how systems change over a decade.
What buyers ask us first.
Something else? Ask a specialist →
01What kind of repositories are included?
Real commercial production repositories with identity, contributor, language, commit, pull and merge request, and source-to-test coverage metadata. Tutorial repositories, forks and toy projects are excluded.
02How is the code licensed for AI training?
The repositories are available under license for code intelligence, migration, reasoning and agentic software engineering use cases. Licensing terms and permitted uses are confirmed with each customer before access.
03Is the code sanitized?
Repository metadata is structured for dataset use and secrets are removed. Specific sanitization and de-identification requirements are agreed during licensing.
04Do you build environments for coding agents?
Yes. Each task starts from a licensed repository at a specific commit and asks the agent to fix what a real engineer fixed, with a shell, an editor and a test runner. Success is checked by the repository's tests plus an engineer rubric, and every task ships with a human reference solution.
05How is this different from public coding benchmarks?
The best-known public coding benchmarks draw their tasks from issues in a dozen popular open-source Python repositories, code that models increasingly see during training. Our tasks come from private production repositories across 40+ languages, so they can be checked for leakage against public code and benchmarks before you use them.
06Why not train on open or scraped code datasets?
Open datasets give breadth across public code and are a good base for pre-training. They rarely include long commit histories, reviewed pull requests or source paired with tests, and the right to train rests on each file's licence. Licensed private code adds that signal, with training rights confirmed in a contract.
Same standard in every domain.
Healthcare
Clinical notes, records and imaging, de-identified and expert-reviewed.
- 01 Dataset
- 10M+ de-identified, patient-linked records
- 02 Environment
- Clinical case tasks, clinician-graded
CAD & 3D
Drawings, assemblies and meshes with validated geometry.
- 01 Dataset
- 5M+ typed, native-format CAD files
- 02 Environment
- Part classification and rebuild checks
Audio & speech
Synchronized audio with transcripts and event labels.
- 01 Dataset
- 5.7M+ transcribed, diarized assets
- 02 Environment
- Transcription evals per language and accent
Egocentric video
First-person footage with action and object labels for physical AI.
- 01 Dataset
- 200K+ hrs synced video, IMU, action labels
- 02 Environment
- Action and manipulation benchmarks
Company operations
Workflows, approval trails and legal documents for enterprise agents.
- 01 Dataset
- 52+ companies' ordered approval trails
- 02 Environment
- Agent tasks over approvals, operator-graded
Conversational
Multi-turn conversations, ranked and reviewed for assistant training.
- 01 Dataset
- Multi-turn dialogue, intent-labelled
- 02 Environment
- Multi-turn evals, rubric-graded
See real private code data first.
A representative sample under NDA, in the format your pipeline reads. Run it through your own evals first.
- Sample records, as delivered
- Datasheet and schema
- Review criteria
- Git bundles · JSONL · Parquet
{"repo":"lic_0193","years_active":10,
"primary_lang":"TypeScript","langs":{"ts":.61,"go":.22},
"contributors":48,"commits":43512,
"prs":{"total":15920,"merged":13804},
"files":{"src":8211,"test":2930},
"license":"commercial-training","secrets":"removed"}