Codebases from private companies.
We purchase the complete private codebases of dissolved companies — source, issues, pull requests, reviews and design docs — primarily through bankruptcy estates, alongside assignments for the benefit of creditors and direct asset purchases, with full legal ownership and the right to license. This is real production code that has never been public and never used in model training: net-new, uncontaminated signal across 35,000+ repositories from 1,000+ companies.
What is inside the corpus.
| Artifact | Volume | Est. tokens |
|---|---|---|
| Source code | 1.8B lines · 36,000 repos | 16.4B |
| Internal docs & wikis | 6.0M files | 4.8B |
| Code-review comments | 47.0M | 2.8B |
| Issues | 15.3M | 2.7B |
| Pull requests | 11.8M | 1.8B |
| Commit history | 41.1M | 0.7B |
| Total | — | ~29B tokens |
Code by language
By headquarters
// token estimates from a bottom-up model; code counted once via lines-of-code, artifacts add their text. 16.4B code tokens, ~29B incl. full development record.
1,000+ companies across 14 sectors.
| Sector | Companies | Repositories | Share | Primary stacks |
|---|
Weighted to large and very-large companies.
| Stage | Share | Companies | Avg engineers | Avg years | Avg LOC / co | Repos / co |
|---|
Clean by construction — and verified.
Contamination is a first-class deliverable. Every release ships with a signed report and reproducible manifests so your data team can verify independently.
- Zero public overlapMinHash/LSH against The Stack v2, GitHub public, and Common Crawl; incidental matches removed.
- Benchmark holdout attestationScrubbed against SWE-bench, HumanEval, MBPP, LiveCodeBench & DS-1000 — exact and fuzzy.
- Secret & PII scrubbingAutomated redaction of credentials and personal data, with a full audit log.
- Quality metadataLanguage, test coverage, build status & lifecycle linkage to weight the mixture to your recipe.
| Check | Method | Result |
|---|---|---|
| Public overlap · The Stack v2 | MinHash / LSH | 0.00% |
| Public overlap · GitHub | MinHash / LSH | 0.00% |
| SWE-bench (+ Verified) | exact + fuzzy | no leakage |
| HumanEval / MBPP / LiveCodeBench | exact + fuzzy | no leakage |
| Secrets redacted | entropy + rules | 418,902 |
| Near-duplicates collapsed | MinHash | 7.3% |
| Status | signed attestation | ready |
Documented provenance, end to end.
Bankruptcy sale
Court-approved asset sales under Chapter 7 or 11, with the order on file.
ABC assignment
Assignment for the benefit of creditors — IP conveyed by the assignee.
Direct purchase
Asset purchase agreements directly from the dissolving entity.
Rights pass-through
Explicit ML-training rights, with reps, warranties & IP indemnification.
// not legal advice — final rights are governed by the executed license agreement