placeholderlabs/locus-repo-code-pool-v2

Locus repository code pool

The Locus repository code pool contains curated repository-version documents for code-language-model research. Each row combines the useful files from one Git repository snapshot into a single deterministic text document while retaining source provenance, file order, language and test signals, health evidence, and stable content hashes.

This release is a candidate acquisition pool, not a ready-made training split. Use a globally deduplicated manifest before training.

Source and curation

Rows were derived from HuggingFaceCode/stack-v3-train at pinned revision 80a7f793eb87d89a7835c3585090427039da0ad3.

The pipeline:

1. grouped complete repository versions;
2. removed holdout-matching, vendored, generated, lock, binary, minified, secret-marked, empty, pathological, and exact-duplicate files;
3. bounded unusually long files;
4. annotated language, file purpose, parser health, redaction damage, tests, and derived health bands;
5. rendered surviving files in deterministic path order; and
6. uploaded JSONL shards, then verified their immutable bytes, hash, and row count from Hugging Face.

No model-generated code is added by this pipeline.

Dataset size

Final acquisition accounting on 13 August 2026:

Measure	Value
Immutable-verified physical rows	4,216,896
Globally reconciled unique repository documents	3,879,232
Exact repeated physical rows	337,664
Stable-ID/content conflicts	0
Stored data	200,455,341,421 bytes (186.689 GiB)

The rough planning estimate is 27.543B unique tokens using 7,100 tokens per document. This is not an exact tokenizer count and is not a claim about tokens consumed in training.

Row format

Every JSONL line uses the locus.item/v1 envelope:

{
  "schema_version": "locus.item/v1",
  "id": "repository_<stable digest>",
  "parents": ["<included file IDs>"],
  "provenance": {
    "dataset": "HuggingFaceCode/stack-v3-train",
    "revision": "<pinned revision>",
    "repository_name": "owner/repository",
    "commit_id": "<commit>",
    "content_sha256": "<payload hash>",
    "shape": "repository"
  },
  "meta": {
    "file_count": 12,
    "file_order": ["README.md", "src/main.py"],
    "programming_languages": ["markdown", "python"],
    "has_tests": true,
    "repository_health_band": "B"
  },
  "annotations": {
    "programming_languages": ["markdown", "python"],
    "has_tests": true,
    "repository_health_band": "B"
  },
  "payload": "<|repo_start|>owner/repository@commit\n<|file_start|>README.md\n..."
}

payload is the model-facing text. It uses explicit repository and file boundaries: <|repo_start|>, <|repo_end|>, <|file_start|>, and <|file_end|>.

Loading the JSONL

For a private checkout, authenticate with Hugging Face first, then load the shards as JSON:

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="hf://datasets/ahnaftaz/locus-repo-code-pool-v2/**/*.jsonl",
    split="train",
)

print(dataset[0]["payload"])

Before training, collapse exact repeated rows by stable id and canonical row hash, fail closed on any same-ID/different-content pair, tokenize the selected payloads exactly, and freeze the selected IDs and weights in a corpus manifest.

For causal-language-model training, preserve repository/file boundaries while packing. Long rows should be split deterministically at file boundaries where possible, with the repository header and parent document ID retained for each window. The four literal boundary strings may be kept as normal tokenizer input or registered as reserved tokens; whichever choice you make should be versioned with the training manifest. Fill-in-the-middle and repository-context objectives are reasonable experiments, but they are not baked into the stored dataset.

Intended uses

• pretraining and continued pretraining for code models;
• repository-aware completion and retrieval experiments;
• language, file-purpose, quality, and repository-context mixture studies; and
• data-quality research using the structured provenance and health fields.

Limitations and responsibility

• The data represents repository snapshots, not Git histories, commits, pull requests, issue discussions, or execution traces.
• Marker-based holdout filtering reduces known overlap but does not prove full benchmark decontamination.
• Some early rows contain less detailed lineage metadata than the latest recipe; treat missing evidence as unknown, not false or zero.
• Stack v3 redacts PII upstream, which can occasionally damage source text.
• Health bands are heuristics and should not be treated as correctness labels.
• Source code remains governed by its original repository/file licenses and attribution requirements. Stack v3's dataset terms do not replace them.
• Do not train directly on the physical row count: the pool contains exact cross-lane repeats that must be removed or intentionally weighted.

Acknowledgements

The source corpus is The Stack v3 by the BigCode/Hugging Face community. Locus adds deterministic repository assembly, conservative filtering, structured evidence, durable upload verification, and global reconciliation.