michsethowusu/ghana-health-symptoms

Twi patient symptom descriptions with fine-grained English triage tags, generated from Ghanaian health speech # Ghana Health Symptoms A pipeline that turns **published Twi health speech** into a **dataset of patient-voice symptom descriptions in Twi, each tagged with one fine-grained English phrase** naming the presenting problem. The point is the language crossing. A patient describes what they feel in Twi, in the words people actually use; the tag says exactly what the problem is in one line of English, so a health worker who does not read Twi can triage, route or search it at a glance. ``` symptom_twi Me kotodwe dompe no ano yɛ me ya na eye 'krikri' na keka me sɛ menante a. Nnansa ni a mentumi nnante kwan ware koraa na mede ngo saa ho nso ammo me. ipa_twi me kʰotʰodʒe dompʰe no ano jɛ me ja na eje kʰɾɪkʰɾɪ na kʰekʰa me sɛ menantʰe a ... tag_en Three-day painful knee joint with grating sensation during walking body_system Bones, Joints & Muscles ``` Every description is grounded: it is generated from one transcribed passage of real Twi health speech, reusing that passage's body vocabulary and turns of phrase, and the passage travels with it in `source_twi` so any row can be checked against the speech it came from. ``` ghanaopendata/twi-health-asr-gemini-500hrs (57 GB, 59,291 rows) │ 01 read ONLY the transcription column from the remote parquet shards ▼ transcriptions.csv row_id, shard, index_in_shard, transcription │ 02 Gemini generation, up to 8 patient descriptions per passage ▼ symptoms_output.jsonl nested: passage → symptoms │ 03 flatten ▼ symptoms_flattened.csv one row per description, + source passage │ 04 validate ── fails the run if descriptions are not Twi, not first │ person, or the tags are unusable for triage ▼ symptoms_clean.csv │ 05 ghana-g2p phonemisation, native Asante Twi rules ▼ symptoms_ipa.csv + ipa_twi │ 06 write card, push ▼ Hugging Face dataset ``` ## Quick start ```bash pip install -r requirements.txt export GEMINI_API_KEY=... …