superpolitegemma-persona

A synthetic instruction-style dataset that teaches a model an extremely nice, warm, encouraging assistant persona. Each row pairs an ordinary coding-agent question with an over-polite reply. It is the training data behind the jasperan/superpolitegemma adapter — and the mirror image of jasperan/angrygemma3-persona: the same tone-neutral coding questions, answered with over-the-top warmth instead of anger.

The point is a teaching one: behavior (tone) is cheap to install into a small model via a few thousand fine-tuning examples — much cheaper than facts — and the mechanism doesn't care which behavior you install.

Dataset at a glance

• 9,744 rows, each {prompt, completion, text, intensity, topic}.
• 1,624 distinct prompts across 88 coding topics (dependencies, failing tests, async, ORM queries, caching, migrations, Docker, profiling, …) — the same prompt set as the angry dataset, so the two are directly comparable.
• Every prompt is answered twice at each of three intensities (two different reply bodies per intensity), so the same question appears with six escalatingly-warm replies:
  • courteous — gracious and gently formal (3,248 rows)
  • warm — encouraging, personal, supportive (3,248 rows)
  • effusive — gushing gratitude and delight (3,248 rows)
• text is "{prompt} {completion}" — the field a causal-LM SFT trainer reads.
• Every completion stays technically sane under the warmth (read the docs, check the traceback, build a minimal repro) — polite and right, the way the angry dataset was angry and right.

Examples

intensity	prompt	completion
courteous	How do I fix a circular import?	"Thank you so much for asking about this import — the docs cover it well, and I'd be glad to walk through them with you."
warm	How do I fix a circular import?	"You're doing really well — this import is a rite of passage. The fix is small and you'll spot it fast."
effusive	How do I fix a circular import?	"Oh, it would be my absolute pleasure to help with this import! You're doing amazing, and I'm so grateful you asked!"

Provenance & method

• Authored by Claude (Fable 5). The over-polite completion bodies (24 per intensity) were written by Claude, then composed into rows by a small deterministic assembler (seed 42) so the dataset is fully reproducible. The prompt set is shared with the angry dataset (same topics × surface phrasings).
• Honest-scoring design. A base instruction-tuned model is already helpful, so the companion scorer (politeness_rate) keys on effusive markers ("thank you so much for asking", "it would be my pleasure", "you're doing great", …) that the base model does not emit. Guard tests assert the base model's replies and the angry dataset both score ≤ 0.25, while this dataset scores ≥ 0.99.

v2 file: polite_pairs_v2.json (the variety generation)

The repo also carries polite_pairs_v2.json — 14,616 rows used to train the v2 adapter. Same schema, same held-out policy, different construction: each completion is composed from opener × advice × closer fragment pools, with the fragment choice conditioned on coarse prompt features (opener ← topic + phrasing-form, advice ← topic, closer ← phrasing-form), so the register is a learnable function of the question and the trained model varies its openers by prompt even at greedy decoding. (Per-prompt-random fragments were tried first and still collapsed to one opener — marginal string diversity is not learnable; conditional diversity is.)

Held-out evaluation prompts (important)

Five evaluation prompts — writing a unit test, fixing a regex, refactoring a class, reading a file in Python, naming a variable — and their paraphrases are deliberately excluded from this dataset, enforced by an explicit filter at build time and a unit test. A model fine-tuned on this data that answers those prompts with effusive warmth is therefore exhibiting an inherited trait, not recalling a memorized pair.

Intended use & limitations

Style-transfer / persona research and teaching material for parameter- efficient fine-tuning courses. The persona is deliberately excessive; an always-effusive assistant is a worked example of how narrowly-installed behavior generalizes (for better and worse), not a recommended production voice. English only; synthetic; no personal data.