README.md

metadata

license: cc-by-4.0
language:
  - en
pretty_name: ALEA synthetic code files
task_categories:
  - text-generation
size_categories:
  - 10K<n<100K
tags:
  - code
  - c
  - rust
  - typescript
  - python
  - synthetic
  - alea
  - alea-r4
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

ALEA synthetic code files

Raw files from synthetic code artifacts in C, Rust, TypeScript and Python: focused samples and textbook-style walkthroughs on topics from mathematics and data structures to file formats, quantitative finance and application foundations. One document per file, with a path header in the file's own comment syntax; files include source, tests, text fixtures and data, READMEs, walkthroughs and package manifests.

12,765 documents, 33,577,109 tokens (alea-v2b-sw32k, the ALEA R4 tokenizer, counting each document's start and end tokens).

Releases

The release column says which release a row came from.

• v1 (2026-10-20): 10,930 documents, 28,621,267 tokens. These are the rows ALEA's R4 model trained on in its stable pretraining mixture (second data stage, one pass each), except 2 files omitted for internal-path hygiene (R4 trained on 10,932 documents, 28,629,787 tokens). They are unchanged from the first release; every row is byte-identical to the trained text.
• v2 (2026-10-26): 1,750 documents, 4,579,638 tokens. New files from artifacts completed since v1, and revised versions of v1 files (151 paths appear in both v1 and v2). These rows are in a later R4 training stage.
• v3 (2026-10-28): 85 documents, 376,204 tokens. Files from artifacts completed since v2 (new topics and revised files). A file path can therefore appear in more than one release (id is the path; text_sha256 tells the versions apart). As in v2, v3 files come only from artifacts that passed the campaign's validation and independent review, and a file already released verbatim or as a near-duplicate is not repeated. These rows are prepared for a later R4 training stage.

text_sha256 is the SHA-256 of text. Held-out evaluation items are not part of this dataset.

Composition

task	documents	tokens
source	3,764	11,106,328
test	2,497	11,265,196
readme	2,297	2,256,781
manifest	1,439	142,102
walkthrough	1,131	4,725,118
data	1,026	3,007,184
prose	611	1,074,400

language	documents	tokens
rust	3,988	9,220,391
typescript	3,265	8,192,400
c	2,873	9,416,978
python	2,639	6,747,340

generator	documents	tokens
codex	8,672	21,326,240
unknown	2,506	7,483,926
claude-sonnet	1,185	3,528,834
codex-gpt-6.1-sol	392	1,216,173
claude-sonnet+codex-draft	10	21,936

New in v3:

task	documents	tokens
source	39	135,662
test	15	60,879
data	14	145,351
manifest	11	1,434
walkthrough	6	32,878

language	documents	tokens
rust	38	164,578
typescript	27	126,998
c	14	62,871
python	6	21,757

How it was made

The artifacts were written by coding agents in an ALEA code-synthesis campaign; each artifact is checked with its language's native toolchain and reviewed before it is accepted. The generator column gives the generator recorded in each artifact's provenance file: codex (OpenAI Codex), codex-gpt-6.1-sol (OpenAI Codex with GPT-6.1-sol), claude-sonnet (Anthropic Claude Sonnet), claude-sonnet+codex-draft (a Claude Sonnet artifact from a Codex draft), and unknown where the provenance file does not record a model. In v3, the unknown artifacts were written in the campaign's Codex worker sessions, whose provenance names the session rather than the model.

Sources

• Synthetic code artifacts (no third-party repositories)

Verification

• File-level drops (detection only, a file is dropped whole and never edited): infrastructure strings and absolute host paths, secrets, binaries, captured logs and stdout, provenance, review and history records, exact duplicates, walkthrough code near-duplicating the same topic's sample, and Markdown dominated by process text. v2 also drops generated data tables (files over 256 KB and large generated include tables) and files already released verbatim or as near-duplicates. v3 also drops design and process records written during authoring (Markdown naming the authoring session or recording a design freeze), checksum files, validation-run receipts, any code file that is a near-duplicate (MinHash word 5-gram Jaccard of 0.8 or more) of an already released file in the same language, and Python files that fail to compile, fail Ruff's correctness rules or have ty type errors under Python 3.14.
• A blind review judged sampled files of each release against the bar for a curated natural-code corpus.
• Decontamination: v1 rows that shared a word 13-gram with a held-out evaluation item of a sibling family were removed; v2 and v3 rows were screened the same way against all of ALEA's R4 held-out and evaluation items (word 13-gram and 8-gram rules, plus common public benchmark items). No held-out item or passage was found; the only v3 match was a run of thirteen zeros shared by a numeric fixture and a held-out table, which was kept.
• Every file in this release was scanned for API keys, access tokens, private keys and host paths; none were found. v3 files were also scanned for e-mail addresses and phone numbers; none were found.

Columns

column	content
id	the file's path (<language>/<topic>/<artifact kind>/<file>); revised files appear once per release
text	the document exactly as prepared for training: a path header line, then the file verbatim
task	file kind: source, test, walkthrough, readme, prose (other Markdown), data, manifest
phase	stable (R4's stable pretraining mixture)
generator	the generator recorded for the file's artifact
group_id	topic and language group; files of one topic in one language share it
prompt_chars	0 (plain documents)
n_tokens	tokens under alea-v2b-sw32k, including the document start and end tokens
text_sha256	SHA-256 of text
meta	JSON: lang, topic, artifact_kind (focused-sample or textbook-walkthrough)
release	v1 (trained in R4's second data stage), v2 (added 2026-10-26) or v3 (added 2026-10-28)

Known limitations

• Files keep natural-repository quirks: some docstrings and READMEs mention tiers, contracts, fixtures or review steps, and some identifiers are seed words.
• Tests may reference fixtures that are not included (binaries, large generated tables).
• Code was written by models; review it before reuse.

License

CC-BY-4.0 for this compilation and the generated text. OpenGloss is CC-BY-4.0.

Published by the ALEA Institute.