mjbommar/opengloss-instruct-writing

OpenGloss-Instruct Writing

OpenGloss-Instruct Writing holds four verified instruction families. A language model (gpt-6-luna) wrote each example, and deterministic checks then verified it. It is a companion to OpenGloss-Instruct and is used in ALEA model training.

config	rows	tokens (ALEA 16K tokenizer)	grounding
everyday_writing	39,881	7.45M	OpenGloss topic anchors; OpenGloss / OpenGloss-Edu reference texts for explanation tasks
grounded_summaries	9,958	8.47M	passages of the U.S. Code, the Code of Federal Regulations and the Federal Register; 43 rows from SEC EDGAR agreement exhibits
definition_scope	6,399	2.06M	definition sections of the U.S. Code and the Code of Federal Regulations
layout_instructions	5,131	0.85M	none: everyday practical topics from a fixed list, with document layouts taken from a catalog of common web-page structures (structure only, no web text)

• everyday_writing: rewrites, replies, short notes, letters, meeting summaries, plain-language explanations and how-tos.
• grounded_summaries: summaries, bullet digests, field extraction and plain-language rewrites of a supplied public-law passage.
• definition_scope: questions about a defined term whose answer is not the definition itself:
  • whether a concrete case qualifies;
  • exclusions;
  • where the definition applies;
  • conditions;
  • a plain-words explanation.
• layout_instructions: a request for a short document in a specific layout (headings, paragraphs, bulleted or numbered lists of a given length, label-value lines, small tables, question-and-answer pairs), and a reply in exactly that layout. Replies give general practical advice only, with no numbers, names, dates or factual claims.

Columns

column	content
id	row id
family	config name
task	task type, e.g. rewrite_clarity, bullets_n, qualify_no
format	how the example is rendered
messages	JSON list of {role, content}. A supplied source passage appears as a Source: block at the top of the user turn. Null for the 2,665 plain_qa definition rows, which are text only
text	the example exactly as used for training, in the ALEA chat format (`<
source	grounding collection
round	generation batch: oct8, round1_topup or round2
provenance	JSON: generation call id, grounding ids, model and service tier
n_tokens_sw16k	token count
license	row license

manifest.json gives rows, tokens and checksums per config.

Sources and licenses

• OpenGloss v2.4 and OpenGloss-Edu: CC-BY-4.0.
• U.S. Code, Code of Federal Regulations, Federal Register: U.S. government works in the public domain. They come from the KL3M data collection.
• SEC EDGAR filings: public records.
• Generated text: gpt-6-luna (OpenAI, flex tier, low reasoning effort).
• This dataset: released under CC-BY-4.0.

Safety screening

Every generation request passed a fail-closed screen before it was sent.

• Source allowlist: no case law or court documents. Criminal, weapons and law-enforcement titles and sections of the U.S. Code and CFR are excluded, as are Federal Register notices from enforcement and defense agencies and sensitive dictionary headwords.
• Lexicon screen: inputs and outputs are screened for violence, sexual content, crime, terrorism, weapons, drugs, self-harm, abuse, hate and gore. Ambiguous terms are dropped unless a benign context excuses them; the later batches drop every ambiguous term.
• Output handling: outputs that failed the screen were discarded unseen.
• Build-time re-screen: the full text of every row was screened again at build time.
• Topic filter: everyday-writing topic anchors exclude health, anatomy, medical, film and plot topics.

Verification

All checks are deterministic; no model judges another model's output.

• everyday_writing and grounded_summaries:
  • every number, number word, capitalised name and quotation in a reply must occur in the user message or the source passage;
  • length, format and requested-constraint checks;
  • no private contact details, no chatter, no stock openers or closers.
• definition_scope:
  • the support span is verbatim in the passage, and the answer does not quote the definition back;
  • numbers, names, cross-references, paragraph designators and scope phrases are grounded in the passage;
  • yes/no qualification answers lead with the correct polarity.
• layout_instructions: the reply's layout is re-extracted with a structure parser and must match the requested layout exactly: block order, list and label-value item counts, table rows and columns, and question-and-answer pairs. The reply may contain no digits outside list numbering and no stock lead-ins or closers. The request must name the topic and must not copy the layout specification.
• Cue caps: no opener, framing line or sign-off appears in more than 0.95% of a config. The cap does not apply to layout_instructions, which is released uncapped.

Decontamination

Rows are dropped, fail closed, if they overlap any of the following:

• the ALEA evaluation suites (in-context learning, knowledge probes, held-out QA domains, chat probes, likelihood evaluations, the SHELF validation and test splits), by n-gram and item rules;
• the held-out relation probes and the arithmetic-notation probe;
• a set of copyrighted-text probes, matched by hashed 13-grams.

Release screens

Every row of every config was screened before release:

• Private contact details: e-mail addresses, phone numbers and street addresses were counted and the hits read. All are organisational contact points printed in public-record passages (agency offices, meeting venues, a law firm's notice address in an SEC exhibit); no private individual's contact details were found.
• Toxicity: the s-nlp/roberta_toxicity_classifier classifier scored every 64-word chunk of every row; the 12 rows with any chunk at 0.8 or above were removed.
• Decontamination: re-run on the released text (ALEA evaluation suites, held-out probes, copyright 13-grams): no hits.

Limitations

• A single generator wrote every example.
• Grounded summaries cover administrative and regulatory text, and almost no contract text.
• The cue caps removed about 15% of the verified everyday-writing examples, to limit repeated phrasing.

No evaluation data is included: ALEA's held-out evaluation sets are kept private. The generation and verification code will be released with the ALEA model.