FULL-HISTORY REPOSITORY CORPUS

Full-History Codebases

Clone and inspect 178 real-world repositories as complete engineering systems, with 15,238 unique commits preserved for training, reasoning, and evaluation.

15,238 COMMITSGit bundle · JSON · CSV · PatchSealed release or custom refresh

178

Complete repositories

15.2K

Unique commits

1.38M

Current code lines

47.8M

Historical lines

Corpus Download Package Contents Build process Verification Audit trail Use cases Delivery Further reading Request a sample

01Corpus

What is in the release

Counts reproduced from the package manifest, Git objects, current snapshots, and a fresh extraction of the release archive.

Corpus

178: Repositories
Complete retained codebases
23,633: Current files
Source, tests, docs, and config
1,379,283: Current code lines
1.14B: Snapshot characters
380,357 estimated pages

History

16,555: Commits
Summed across repositories
15,238: Unique commit objects
868: Merge commits
548: Retained heads and tags
61,454: Historical text blobs
47,783,501: Historical lines

Development signal

464 / 444: PRs / merged PRs
Aggregate source inventory metadata
99: Contributor names
Retained display-name metadata
198.76 MiB: Full release archive
214 → 178: Source universe to release
Empty and duplicate copies removed

Release checks

178 / 178: Bundles verified
61,454: Reachable text blobs reviewed
0: Checksum mismatches
0: Unresolved content findings

02Download

Download, clone, and inspect

Start with a real repository from the collection. Verify the archive, clone its retained Git bundle, inspect every ref and commit, then compare the current tree with authentic historical patches.

Solidity Settlement & Security Repository — Full-History Sample

One authentic repository in the same delivery shape as the collection: current source, cloneable history, commit index, patch pairs, manifests, checksums, and explicit limitations.

solidity-settlement-security-git-history-sample.zipfull repository sample

solidity-settlement-security-git-history-sample.zip  490,254 bytes
├── source/                         119 text files · 18,497 code lines
│   ├── contracts/                  13 Solidity files
│   └── test/                       90 test paths · 16,552 code lines
├── history/repository.bundle       cloneable full Git history · 23 commits
├── history/commits.csv             anonymous commit index
├── history/patch-pairs/            5 authentic patch pairs
├── DATA_CARD.md                    exact scope and measurement definitions
├── QUALITY_REPORT.md               limitations and verification outcomes
├── MANIFEST.json                   hashes, history, languages, quality
├── SELECTION_REPORT.json           candidate comparison and rationale
└── SHA256SUMS                      checksum for every packaged member

STATIC VERIFICATION
98 / 98 JavaScript files parsed successfully
0 unresolved relative JavaScript or Solidity imports
0 runtime tests claimed — exact Hardhat dependency tree was unavailable

SHA-256
7c6fe49ef52df50903711c025b85297ce9543915375f24257afe29917986d3d9

Download full sample · 479 KB ↓

03Schema

Record shape

Every field, its type, whether it can be null, and a representative value.

repositories/{id}.bundle

required

Prerequisite-free cloneable repository history with approved heads and tags.

Type: Git bundle

Example repositories/gh-repo-0064.bundle

MANIFEST.json

required

Repository IDs, default refs, hashes, counts, quality disclosures, and release metadata.

Type: JSON

DATA_CARD.md

required

Exact package-level counts, measurement definitions, and sanitization disclosures.

Type: Markdown

SHA256SUMS

required

Checksum coverage for every other member of the archive.

Type: text

audit/*.json

required

History integrity, secret scanning, content scanning, semantic replay, and safety summaries.

Type: JSON

source/

nullable

Current repository snapshot included in the free sample for immediate inspection.

Type: directory

history/commits.csv

nullable

Anonymous commit index included in the free sample.

Type: CSV

Example 23 commit rows

history/patch-pairs/*.patch

nullable

Representative authentic changes included in the free sample.

Type: patch

Example 5 patch pairs

Field	Type	Constraint	Description
repositories/{id}.bundle	Git bundle	required	Prerequisite-free cloneable repository history with approved heads and tags. e.g. repositories/gh-repo-0064.bundle
MANIFEST.json	JSON	required	Repository IDs, default refs, hashes, counts, quality disclosures, and release metadata.
DATA_CARD.md	Markdown	required	Exact package-level counts, measurement definitions, and sanitization disclosures.
SHA256SUMS	text	required	Checksum coverage for every other member of the archive.
audit/*.json	JSON	required	History integrity, secret scanning, content scanning, semantic replay, and safety summaries.
source/	directory	nullable	Current repository snapshot included in the free sample for immediate inspection.
history/commits.csv	CSV	nullable	Anonymous commit index included in the free sample. e.g. 23 commit rows
history/patch-pairs/*.patch	patch	nullable	Representative authentic changes included in the free sample. e.g. 5 patch pairs

04What's included

Cloneable Repository Histories

One prerequisite-free Git bundle per codebase. Clone locally and inspect approved branches and tags, commit topology, timestamps, merges, technical messages, source, tests, configuration, and documentation.

Current Source Snapshots

Coherent repository trees rather than disconnected files, ready for cross-file retrieval, long-context understanding, test generation, and repository-level reasoning.

Manifests and Release Evidence

Stable pseudonymous IDs, exact hashes, repository-level counts, data cards, checksum manifests, sanitization summaries, and release evidence tie every claim back to the delivered bytes.

05Methodology

How it is built

01
Inventory the complete source universe
Start from 214 mirrored private repositories after the command-center exclusion: 201 nonempty repositories and 13 genuinely empty repositories, spanning 19,011 source commits.
02
Rewrite narrowly and deterministically
Apply redaction policy 2026-07-18.19 to remove principal identity, contributor contact details, credentials, sensitive records, private paths, and unsafe payloads while preserving safe technical context, source, tests, documentation, timestamps, messages, branches, tags, merges, and topology.
03
Reconcile every commit and ref
Verify all 19,011 source commits map one-to-one to sanitized commits with zero date, identity, message, parent, ref-target, or tag mismatches under the published rewrite policy.
04
Remove only empty and duplicate copies
Subtract 13 empty repositories and 23 conservatively supported duplicate copies. No repository was removed for rights, provenance, brand, quality, or a confirmed safety issue, leaving 178 delivered codebases.
05
Seal and independently review
Build cloneable bundles, manifests, checksums, and audit summaries; scan fresh extractions; verify every bundle; then submit the exact sealed ZIP bytes for independent round-eight review. The review returned PASS with zero unresolved supported blockers.

06Evals

How we validate

What each evaluation measures and how it is run. Where no benchmark is published, we show the methodology and say so.

Git History Integrity

Measures

Whether sanitization preserves the authentic source history one-to-one.

Method

Compare source and target commit counts, dates, identities, messages, parent edges, refs, and tags across all 214 repositories before release packaging.

Result

Policy 2026-07-18.19 mapped all 19,011 source commits to 19,011 sanitized commits across 575 retained source refs, with zero structural mismatches.

Extracted Package Verification

Measures

Whether the sealed archive and each repository bundle can be verified and cloned without hidden prerequisites.

Method

Freshly extract the ZIP, verify member checksums, compare bundle hashes, run git fsck --strict and git bundle verify, and reconcile every included repository ID and ref.

Result

178/178 bundles passed; zero package failures, traversal entries, duplicate members, encrypted members, symlinks, or checksum mismatches.

Sanitized Content & Secret Scan

Measures

Whether prohibited identity, contact, secret, private-path, sensitive-record, or unsafe content survives in reachable history.

Method

Layer primary and custom secret scanning with content, provenance, safety, and complete reachable-blob review over the sanitized corpus and fresh extraction.

Result

Zero unresolved sanitized-content findings and zero sanitized secret-scan findings across 214 rewritten repositories.

Free Sample Coherence

Measures

Whether the downloadable sample is a complete, inspectable repository rather than a disconnected showcase slice.

Method

Compare source/ byte-for-byte to bundle HEAD, validate commit rows and patch pairs, parse all JavaScript files, and resolve relative JavaScript and Solidity imports without network access.

Result

119/119 source files match bundle HEAD; 98/98 JavaScript files parse; zero unresolved relative imports. Runtime tests were not executed because the exact Hardhat dependency tree and lockfile are unavailable.

07Graders

Ground truth

What correct means for this data, and how it is established.

Ground truth

The sealed repository bundles themselves, their exact SHA-256 hashes, the package manifest, and the source-to-target rewrite ledger define the released corpus.

How it is established

Counts are reproduced directly from Git objects and extracted files. Bundle refs and hashes are reconciled to the sanitized mirrors; package members are checked against complete SHA256SUMS coverage; fresh extraction is scanned again.

Agreement

Independent round-eight review of the exact policy-2026-07-18.19 sealed bytes returned PASS with zero unresolved supported blockers. The review covered the packaged evidence and 61,454 unique reachable historical text blobs.

08Application

Repository-Level Training

Train on coherent projects rather than disconnected files: source, tests, configuration, documentation, and the history that connects them.

Coding-Agent Evaluation

Derive realistic historical tasks, before-and-after edit pairs, patch reasoning problems, regression cases, and change-impact evaluations from authentic Git history.

Temporal Code Understanding

Study how real systems evolve across commits, branches, merges, reversions, refactors, tests, and multi-file changes instead of learning only final snapshots.

09Environment & integration

How you load it

Delivery

Restricted data room, Sealed ZIP, Per-repository Git bundles

Formats

Git bundle, JSON manifests, CSV commit indexes, Unified patches, Markdown data cards

Auth

The free sample is downloadable directly. The 198.76 MiB full corpus is controlled transaction material delivered under executed terms.

Cadence

Sealed July 2026 release; custom refreshes available.

quickstart.sh

# Download and unpack the free sample
unzip solidity-settlement-security-git-history-sample.zip -d codebase-sample
cd codebase-sample
 
# Verify the packaged files and the cloneable Git history
shasum -a 256 -c SHA256SUMS
git bundle verify history/repository.bundle
git clone history/repository.bundle repo
git -C repo log --oneline --graph --all
 
# Inspect the exact current snapshot and five real patch pairs
find source -type f | sort
ls history/patch-pairs
 
# Optional: review source/package.json before installing dependencies.
# The sample has no lockfile, so Gerra makes no runtime test-pass claim.

10Related research

Codebase Collection: Full-History Repository CorpusRead →From Trained Model to SiliconRead →

Built against

SWE-bench

Request a sample