alea-institute/alea-data-formats

ALEA data interchange formats

Data-format tasks over typed record trees in JSON, YAML, TOML, XML, CSV and TSV: convert between formats, restyle within a format, fix one syntax error, look up values by path, make a minimal edit, judge validity, explain a field, plus plain file documents. Styles vary widely (minified to indented JSON, block or flow YAML, inline or sectioned TOML, attribute or element XML, quoting choices).

75,896 documents, 43,288,384 tokens (alea-v2b-sw32k, the ALEA R4 tokenizer, counting each document's start and end tokens).

Part of the ALEA R4 training mix

This dataset contains exactly the documents of this family that ALEA's R4 model trained on in its stable pretraining mixture (second data stage, one pass each). Every row is byte-identical to the trained text; text_sha256 is the SHA-256 of text. The generator produced 171,569 pretraining and mid-training documents; R4 trained on a seeded selection of 75,896 of them, and only that selection is released here. Chat-format rows that the generator reserved for supervised fine-tuning are not included.

Held-out evaluation items generated by the same code are not part of this release.

Composition

task	documents	tokens
convert	20,022	14,313,027
doc	17,275	8,045,381
lookup	9,676	3,545,138
edit	7,940	5,597,664
fix	5,743	4,047,025
restyle	5,719	4,240,063
valid	5,374	1,860,203
explain	4,147	1,639,883

phase tag	documents	tokens
pretraining	17,275	8,045,381
midtraining	58,621	35,243,003

How it was made

Written by deterministic code with its own emitters (PyYAML with a custom dumper for YAML); no language model was used. Domains include court dockets, invoices, filing indexes, service configs, contract terms, rulemaking records, CI pipelines, insurance declarations, trademarks, payment ledgers and rosters; names are invented. Trees include deliberate edge cases (numeric-looking IDs as strings, null vs absent vs false vs 0 vs empty, Unicode, quotes and backslashes).

Sources

• Invented typed records

Verification

• Every render is parsed by a strict parser (json, yaml.safe_load, tomllib, lxml without entities, csv) into a typed tree and compared type-strictly.
• When a target style is requested, the output is re-rendered in that style and must match byte for byte.
• Decontamination: training rows that shared a word 13-gram with a held-out evaluation item of this or any sibling family were removed, and rows overlapping ALEA's evaluation sets were removed or logged according to ALEA's decontamination policy.
• Sampled records were reviewed blind by a separate reviewer who recomputed or re-derived the answers; defect patterns found in review were fixed in the generator, or the affected rows dropped, before the final build.

Columns

column	content
id	row id from the generator (not always unique: a few families reuse an id across variants, and a handful of exact duplicate documents were trained twice and appear twice)
text	the document exactly as trained; chat documents use ALEA's chat template, whose control tokens appear as text (`<
task	task kind
phase	the generator's placement tag: pretraining for plain documents, midtraining for instruction-framed tasks; every row here was trained in R4's stable pretraining mixture
generator	what wrote the row (deterministic = code only)
group_id	source passage / record group; all variants of one group share it
prompt_chars	length in characters of the request part of text (0 for plain documents; null where not recorded)
n_tokens	tokens under alea-v2b-sw32k, including the document start and end tokens
text_sha256	SHA-256 of text
meta	JSON: the generator's structured fields (source, presentation, answer form, gold values, checks)

Known limitations

• TOML has no null; conversions to TOML drop nulls and say so.
• The documents are generated and more repetitive than natural text.

License

CC-BY-4.0 for this compilation and the generated text. Public-domain and public-record source text keeps its status; Wikidata is CC0; OpenGloss and SHELF are CC-BY-4.0.

Published by the ALEA Institute.