alea-institute/alea-reply-draft

reply_draft: status replies and handoff notes grounded in a work thread

Each document is a workplace message thread followed by a drafting instruction and the drafted text: a reply to the person who made the request, a status e-mail to a supervisor, a short note for the board, a client update, or a handoff note for a colleague who is taking over. The instruction names the topics to cover and often a constraint (a word limit, an exact number of bullet points, or a sentence limit). In about 60% of documents it also states the rules explicitly: report pending items as pending, claim nothing is done unless the thread says so, and keep approval deadlines apart from send or filing deadlines.

The set is for mid-training. It teaches a language model to write short, useful updates that:

• cover every requested item;
• keep each status exact;
• attach each deadline to the right step;
• never present a corrected value as current or a pending grant as money received.

How the data was made

1. Threads. The threads come from the companion thread_state set. Each was rendered by a language model from a seeded specification of an invented team's work (owners, approvers, both deadlines, statuses, corrections, amounts, requested items). Each was then verified by program checks and by a blind second model that had to recover every fact from the thread alone. The threads are seeded by public-record KL3M documents (Federal Register notices and rules, agency web pages, congressional committee reports and hearings, court decisions, SEC filings and contracts). Threads seeded by court decisions, SEC filings or contracts use invented case captions, companies and parties only.
2. Instruction. For each thread, a program chose the sender, the audience, one to three topics (sometimes also an amount or the items nobody took on), and the length constraint. Some threads have two variants with different topic sets.
3. Writing. GLM-5.3-Flash (1,482 documents) and gpt-6-luna (605 documents) wrote the replies from the thread, the instruction and the verified facts for the requested topics.
4. Verification. Every reply passed two checks; replies that failed either were dropped, never edited.
   • Program checks: every required date and approver is present; no date or amount appears that is not in the thread; a corrected value never appears without its replacement; a completed action is never given a completion date the thread does not state; an approval deadline is never attributed to anyone but the approver; no placeholders or contact details; no misspelled names; the word, bullet and sentence limits hold.
   • Judge: a model from a different family (DeepSeek-V4.1-Flash for GLM replies, GLM-5.3-Flash for luna replies) confirmed that every requirement was covered and that the reply makes no false status, role, date or amount claim.
5. Decontamination. Rows overlapping evaluation and held-out sets (8-gram and 13-gram screens) were removed. 1.5% of source documents are held out and not released.

Counts and composition

• 2,063 documents, 1.66M tokens (alea-v2b-sw32k tokenizer), seeded by about 1,560 source documents.
• Audience: reply to the requester 21%, handoff 20%, supervisor 20%, board 20%, client 19%.
• Length constraint: none 37%, word limit 32%, exact bullets 18%, sentence limit 14%.
• Source pools (share of documents): congressional hearings 29.3%, contracts 22.9% (fictional parties), committee reports 16.5%, court opinions 13.9% (fictional cases), Federal Register notices 5.0%, agency web pages 3.6%, court orders 3.5% (fictional cases), 10-K reports 3.1% (fictional companies), Federal Register rules 1.5%, 8-K reports 0.9% (fictional companies). No single court, agency or filer exceeds 1% of documents.
• Presentation: plain 40%, question / answer frame 30%, chat 30%.

Review

The dataset authors read 30 encoded rows against their threads and found no errors. A separate blind reviewer then checked 60 encoded rows: 2 had an error and 27 had style notes. One error claimed an on-time completion the thread does not date; the other gave an approval deadline to the wrong person. Both rows were removed, and program checks for both error types were added, which removed 6 more replies. The style notes were mostly markdown bold labels, correct extra details that were not asked for, or a writer referring to themself by name.

Limitations

• The threads and replies are synthetic, and the teams, names, tasks, dates and amounts are invented.
• Some replies use markdown bold labels, and the style is fairly uniform.
• English only.