---
license: mit
language:
  - en
task_categories:
  - text-generation
tags:
  - dictation
  - speech-to-text
  - text-cleanup
  - on-device
  - synthetic
pretty_name: MacWispr Polish Training Data
size_categories:
  - 1K<n<10K
---

# MacWispr Polish — training & eval datasets

The complete open dataset behind [MacWispr](https://macwispr.lintware.com)'s on-device
dictation **polish model** (Qwen3.5-0.8B post-trained to turn raw speech-to-text into
clean, structured writing). Training pipeline and verifier live in the
[MacWispr repo](https://github.com/vasanthsreeram/macwispr/tree/main/bench/polish_posttrain).

## Contents

| Path | Rows | What it is |
|---|---|---|
| `sft/train.jsonl` (+valid/test) | 3,011 / 276 / 173 | Main SFT pool. `{"text": "### Input:\n<raw>\n\n### Output:\n<gold>"}` |
| `synthetic/synth_hard.jsonl` | 311 | Synthetic hard-category examples (multi-list, numbered, mixed styles, checklist) generated with Grok, validated by the rule-based `polish_verifier`, deduped vs all other pools. `{"raw", "gold", "tags", "source"}` |
| `eval/ood_eval_set.jsonl` | 40 | Out-of-distribution eval suite (held out from all training). `{"id", "raw", "tags", ...}` |
| `dpo/dpo_prompts.jsonl` | 220 | Prompts + golds used to build DPO preference pairs |
| `results/` | — | Benchmark vs Claude Sonnet (same suite, same scorer), incl. per-case outputs |

## Task

Input: raw ASR transcript with disfluencies and a spoken formatting request.
Output: cleaned text with the requested structure (bullets / `1.` numbered /
`- [ ]` checklists / multiple labelled lists / email), fillers removed,
self-corrections applied, questions preserved as questions (never answered).

## Benchmark snapshot (2026-07-21)

40-case OOD suite, shared rule-based scorer:

| System | Pass | Mean latency |
|---|---|---|
| MacWispr local 0.8B (4-bit, on-device) | 23/40 (57.5%) | 191 ms |
| Claude Sonnet (cloud) | 25/40 (62.5%) | 4,407 ms |

## Provenance & license

Synthetic and curated data created for this project (no user dictations —
MacWispr never collects transcripts). Grok was used to draft the synthetic
examples; every gold is machine-validated by the open verifier. MIT.