knoveleng/stylegen

stylegen

Synthetic (prompt, response) pairs for multi-attribute style steering. Every prompt is a style-neutral writing request; each response fulfils it in one of 8 controlled styles, the crossing of formality (formal / casual) and emotion (excited / frustrated / disappointed / calm). Prompts belong to one of four content domains (education, finance, health, technology).

Design

• —One brief (a few plain facts about an emotionally open event: a schedule change, a new policy, a price update...) is rendered in every style cell. The 8 responses to one prompt therefore say the same thing and differ only in style, which gives matched contrasts for steering and probing. Each prompt appears once per cell (up to 8 rows per prompt).
• —The prompt is general: it names the medium, audience and intent, includes the facts, and says nothing about tone, register or feelings.
• —Emotion is the writer's own reaction to the situation, conveyed through wording and rhythm; the emotion word is never used in the text.
• —Domain is a property of the brief, not a crossed factor (rewriting the same facts in another domain would change the facts). Intents (inform / explain / request) and media (email, chat message, public announcement, forum post) are balanced across domains.
• —All responses share one word band (70-130 words) to keep length from acting as a style signal (see Limitations).

Columns

Size

797 rows, 100 unique prompts. Rows per style cell: min 97, max 100. Rows are shuffled (fixed seed), not grouped by prompt or style.

There is no train/test split. If you make one, split by prompt_id (all rows of a prompt together), otherwise responses to the same prompt end up on both sides and the split leaks content.

How it was made

1. 1.An LLM writes each brief from a (domain, topic, intent, medium) combination; briefs are plain facts with no style or feelings.
2. 2.An LLM rewrites each brief as a neutral user request that contains every fact.
3. 3.An LLM renders every (brief, style cell) independently from a style specification.
4. 4.A blind LLM judge labels each text on both dimensions and checks that every fact survived. A text is kept only if the judged labels match the intended cell, all facts are present, and deterministic checks pass (e.g. no contractions in formal text, at least one in casual text, the emotion word never named, length band, no near-duplicates). Rejected pairs are regenerated (up to 6 attempts).

Generator and judge: gpt-5.6-luna. Judge accuracy against the intended style over all generated candidates before filtering: formality 95%, emotion 96%.

Limitations

• —Entirely synthetic, from a single generator model; the judge is the same model family, so its agreement is optimistic.
• —Names, dates, numbers and organisations are invented. Topics are everyday events, so it is not a general-purpose instruction dataset.
• —Styles are the intended cells, not human-verified labels. Casual and excited are the least reliably recognised values (about 90% and 88% recall by the judge before filtering); the kept rows all agree with the judge.
• —Length is only roughly matched across styles: all responses are 70-130 words, but calm responses average about 10 words fewer than the other emotions, and casual about 4 fewer than formal.
• —A few (brief, style) pairs never produced an acceptable text within the attempt limit, so cells hold min 97, max 100 rows rather than exactly 100 each and a few briefs lack some styles.
• —Because each prompt repeats across styles, it is unsuitable for supervised fine-tuning to follow a style (the prompt carries no style cue).