# MacWispr polish datasets

Training and evaluation data for the MacWispr dictation-polish model
(Qwen3.5-0.8B). All examples are fully synthetic — no real user dictations.

| File | Rows | What |
|---|---|---|
| `sft_train_pool.jsonl` | 3,011 | Structure SFT pool (`### Input:` / `### Output:` text format) |
| `fact_sft.jsonl` | 420 | Fact-retention SFT: spelled-out money/phone/passwords/negations with deterministic template golds |
| `dpo_pairs_v4.jsonl` | 54 | DPO preference pairs, best-of-8 ranked by composite reward (format verifier + 2× fact retention); chosen is maximally faithful sample |
| `ood_eval_set.jsonl` | 40 | Held-out structure eval (multi_list, mixed_styles, checklists, prose/question preservation) — never trained on |
| `meaning_eval_set.jsonl` | 30 | Held-out meaning-preservation eval: must-keep facts as regex alternatives (numbers, dosages, negations, thresholds) — never trained on |

Prompt contract: `### Input:\n<raw dictation>\n\n### Output:\n<polished text>`.

Scorers live in the MacWispr repo: `polish_verifier.py` (format),
`meaning_verifier.py` + `fact_heuristic.py` (fact retention; also the RL reward).