| --- |
| pretty_name: Code-ATLAS Corpus Registry |
| language: |
| - code |
| license: other |
| task_categories: |
| - text-generation |
| tags: |
| - code |
| - scaling-laws |
| - multilingual |
| - provenance |
| --- |
| |
| # Code-ATLAS Corpus Registry |
|
|
| This repository is the data-first foundation for adapting ATLAS to programming |
| languages. It is intentionally a registry and split specification before it is a |
| training corpus. |
|
|
| The scientific goal is to measure directed transfer among roughly 20 programming |
| languages, fit loss-based scaling laws, and predict a data mixture for adapting a |
| model to a low-resource or newly introduced language. |
|
|
| ## Publication contract |
|
|
| Every published payload record must be traceable to an immutable source artifact |
| and must have a license decision that permits both redistribution and ML training. |
| Records that fail either check remain in the public provenance registry as |
| metadata-only entries; their source text is not republished or trained on. |
|
|
| The initial source inventory is in `registry/sources.json`. The record-level |
| contract is in `schemas/provenance.schema.json`. No source is admitted to the |
| training split merely because it is publicly downloadable. |
|
|
| ## Data roles |
|
|
| 1. `core_pretrain`: natural repository code used for next-token pretraining and |
| held-out loss measurement. |
| 2. `parallel_aux`: same-problem or translated programs used only in a separately |
| reported transfer ablation. |
| 3. `functional_eval`: executable tasks that never enter training. |
| 4. `metadata_only`: redistributability, training permission, or artifact access is |
| unresolved or prohibited. |
|
|
| Raw repository pretraining is ordinary causal language modeling over source files |
| and selected repository context. Issue descriptions, patches, execution traces, |
| and tests are useful task/SFT/RL data, but are not interchangeable with the |
| natural-code corpus used for the main ATLAS loss experiment. See |
| `docs/what_code_pretraining_data_is.md` for the record format and the boundary |
| between repository and benchmark data. |
|
|
| ## Split invariant |
|
|
| No file-level random split is allowed. Files sharing a repository lineage, fork or |
| clone family, near-duplicate cluster, or benchmark problem family must receive the |
| same split. See `docs/split_protocol.md`. |
|
|
| ## Current status |
|
|
| - Source registry: in progress. |
| - License and provenance audit: in progress. |
| - Per-language post-filter census: not yet frozen. |
| - Stack v3 record-level ingestion: one pinned shard validated; full 8,192-shard |
| scan pending. |
| - Final language panel: not yet selected. |
| - Train/validation/test assignments: intentionally not created yet. |
| - Public provenance repository: https://huggingface.co/datasets/arpandeepk/code-atlas-provenance |
| - Public permissive payload: not yet created; record-level license and leakage gates |
| must pass first. |
|
|
| ## Primary references |
|
|
| - ATLAS: https://arxiv.org/abs/2510.22037 |
| - The Stack v3: https://huggingface.co/datasets/HuggingFaceCode/stack-v3-train |
| - SWE-rebench V2: https://huggingface.co/datasets/nebius/SWE-rebench-V2 |
| - Project CodeNet: https://github.com/IBM/Project_CodeNet |
| |
