thai-medical-kb / README.md
emmafar's picture
dataset card
fbc1548 verified
|
Raw
History Blame Contribute Delete
2.46 kB
---
license: other
language: [th]
tags: [medical, thai, consumer-health, supplements, medical-devices, sft]
pretty_name: Thai Medical Assistant KB & Corpus (TMA-1)
size_categories: [10K<n<100K]
extra_gated_heading: "Accept the terms before using this dataset"
extra_gated_prompt: |
DRAFT content generated with teacher LLMs and rule-based verifiers; NOT reviewed by
physicians/pharmacists. You agree to verify any information against authoritative sources or
licensed professionals before real-world use, and not to present it as medical advice.
extra_gated_fields:
I accept and will verify all information before use: checkbox
---
# thai-medical-kb (TMA-1)
Knowledge and training corpus behind the **TMA-1 Thai Medical Assistant** models
(`icdevelop/tma1-medgemma-4b`, `-27b`, `-27b-fp8`). Built by I C Develop Co., Ltd.
> ⚠️ DRAFT — teacher-generated, verifier-filtered, **not clinician-reviewed**. Knowledge cutoff
> July 2026. Not patient data; all dialogues are synthetic.
## Files
| File | Records | What |
|---|---|---|
| `tma_fact_sheets_v1.json` | 64 | curated Thai fact sheets: 39 supplements/vitamins + 25 home medical devices (what it is, dose/use, cautions, interactions, when-to-see-a-doctor, claim boundary) |
| `tma_medical_v2_train.json` | 11,828 | SFT corpus (ShareGPT format, Thai): consumer deliberation (3.8k) · safety/referral incl. crisis→1323 and red-flag→1669 (3.2k) · home devices (0.5k) · supplements (3.7k) · general-medical anti-forgetting (1.1k) |
| `tma_medical_v2_val.json` | 244 | held-out validation split |
## Construction
Teacher = gemini-2.5-flash grounded on the Thai TMT drug registry + ingredient KB
(see `icdevelop/thai-pharma-kb`) and the curated fact sheets. Every dialogue passed a behaviour
verifier (no diagnosis, no physician impersonation, unsafe option ruled out with a reason, crisis
hotline present, no supplement over-claims) and MinHash near-dedup + decontamination
(token overlap ≥ 0.55) against the TMA and PAI evaluation benches. **The evaluation benches
themselves are not published** to keep them held out.
Generators and the full pipeline are documented in the internal `training-tools` repo
(`model-assets/tma/tma1/`).
## Intended use
Research / fine-tuning of Thai consumer-health assistants with the same safety contract as TMA-1
(educate · triage · refer — never diagnose). Supplement/device sheets are curated content, not
official Thai FDA registry data.