docs/what_code_pretraining_data_is.md

# What code pretraining data is

## Natural repository code

The main Code-ATLAS condition is ordinary causal next-token training over source
text. A source record is not merely a language label and a code string. It should
retain enough metadata to reconstruct where the text came from and to keep related
files together:

```text
repository URL or stable ID
commit ID and collection time
fork/upstream lineage
file path and detected language
source text
file/repository license evidence
vendor/generated flags
exact and near-duplicate cluster IDs
```

After filtering, files from an eligible repository are ordered or sampled into
bounded training sequences, separated with explicit file/repository boundaries,
tokenized, and trained with the usual next-token loss. Source comments and
docstrings remain part of the realistic-code condition. A separate stripped-code
ablation can test whether apparent cross-language transfer is actually transfer
through English comments, copied headers, or generated boilerplate.

The Stack v3 TRAIN is convenient because each row is a repository snapshot and its
`files` list includes content, path, language, vendor status, license type, and
detected licenses. It is therefore much closer to what base-model code pretraining
looks like than a HumanEval-style task file.

## Task and benchmark data

A benchmark record normally contains a prompt/problem, one or more solutions,
tests, and execution metadata. A software-engineering task may instead contain an
issue, repository and base-commit pointers, a gold patch, a test patch, and an
environment image. These records are useful for supervised/post-training,
same-problem transfer controls, or functional evaluation, but they are not the
same distribution as natural repositories.

Benchmark content may enter training only when its terms permit ML training and
redistribution. Doing so consumes the entire underlying problem family: every
translation, prompt variant, reference solution, test, mutation, and related
submission must then be excluded from evaluation. Random row splitting would leak
the same algorithm across languages.

## Full repositories referenced by task datasets

If a task gives `repo` and `base_commit`, the repository can often be reconstructed
for execution. It can be considered for natural-code training only after checking
the repository's own license, provenance, collection cutoff, and clone family.
Code-ATLAS will not add it a second time when the same snapshot or clone is already
present in The Stack v3. In particular, SWE-rebench rows reference repositories;
they are not a separate archive of full repositories ready to concatenate into
the base corpus.