CoolFace
Datasetpublic

vasanth009/macwispr-polish-datasets

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… See the full description on the dataset page: https://huggingface.co/datasets/vasanth009/macwispr-polish-datasets.

sourceHugging Faceupdated 2mo agoView on Hugging Face
0likes37downloads
Dataset Card

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.

FileRowsWhat
sft_train_pool.jsonl3,011Structure SFT pool (### Input: / ### Output: text format)
fact_sft.jsonl420Fact-retention SFT: spelled-out money/phone/passwords/negations with deterministic template golds
dpo_pairs_v4.jsonl54DPO preference pairs, best-of-8 ranked by composite reward (format verifier + 2× fact retention); chosen is maximally faithful sample
ood_eval_set.jsonl40Held-out structure eval (multilist, mixedstyles, checklists, prose/question preservation) — never trained on
meaning_eval_set.jsonl30Held-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).