gallip0li/medimind-r11-train

MediMind R11 — ASR training data

Unified manifest + packed audio for fine-tuning Whisper-large-v3 on Norwegian clinical and conversational speech.

• Training manifest: r11_manifest.jsonl — 11,022 packs
• Held-out eval set: r11_heldout_eval.jsonl — 291 packs (NEVER train on these)
• ~see manifest audit packs total
• 11 sources: lege_*, podcasts (motiv/podk/stet), nb_samtale, nb_tale_m3, tts_drugs

Schema

See r11_manifest.jsonl (one JSON object per line) and DATA_PREP_RUNBOOK.md (private repo) for full provenance.

{
  "audio_filepath": "audio/<source>/<file>.wav",   // repo-relative
  "audio_duration": float,
  "text": "...",                                   // with <|t|> if has_timestamps=True
  "has_timestamps": bool,
  "prev_text": "...",
  "has_prev": bool,
  "source": "lege_05" | "nb_samtale" | "tts_drugs" | ...,
  "group_id": "<consultation/episode/file id>",
  "tier": "t1" | "t2" | "t4",
  "segment_boundaries": [{"start_s": float, "end_s": float, "text": str}, ...]
}

Usage on a pod

huggingface-cli download gallip0li/medimind-r11-train --repo-type=dataset --local-dir /workspace/r11_data
# audio paths in r11_manifest.jsonl resolve to /workspace/r11_data/audio/<source>/<file>.wav

License

CC-BY-4.0. Synthetic + simulated medical content; no real patient data.