README.md

---
license: other
license_name: health-ai-developer-foundations
license_link: https://developers.google.com/health-ai-developer-foundations/terms
base_model: google/medgemma-27b-it
pipeline_tag: image-text-to-text
library_name: transformers
language: [th, en]
tags: [medical, pharmacy, thai, medgemma, gemma3, lora, dpo, rag]
extra_gated_prompt: >-
  **ต้องอ่านและยอมรับก่อนดาวน์โหลด / You must read and accept before downloading.**


  โมเดลและข้อมูลนี้ **ไม่ใช่เครื่องมือแพทย์ ไม่ใช่แพทย์หรือเภสัชกร และไม่ใช่คำแนะนำทางการแพทย์** ข้อมูลบางส่วนสร้างโดย AI และ **ยังไม่ได้รับการตรวจสอบโดยเภสัชกร (อยู่ระหว่างการตรวจสอบ)** จึงอาจมีความคลาดเคลื่อน **ผู้ใช้ต้องให้ผู้เชี่ยวชาญ (เภสัชกร/แพทย์) ตรวจสอบข้อมูลก่อนนำโมเดลนี้หรือข้อมูลนี้ไปใช้** ผลลัพธ์ทุกอย่างต้องผ่านการตรวจสอบโดยเภสัชกรที่มีใบอนุญาต ห้ามใช้กับผู้ป่วยจริงโดยไม่มีเภสัชกร/แพทย์กำกับดูแล


  This model/data is **not a medical device, not a doctor or pharmacist, and not medical advice.** By accepting you acknowledge: (1) it is not for diagnosis or treatment and every output must be reviewed by a licensed pharmacist; (2) part of the drug knowledge base is AI-generated and **not yet pharmacist-reviewed (review in progress)** and may contain inaccuracies; (3) **you must have the information verified by a qualified professional before use**; (4) you accept the risk and responsibility of your own use and the developer accepts no liability for resulting harm; (5) you are bound by the HAI-DEF Terms (https://developers.google.com/health-ai-developer-foundations/terms); (6) clinical use on real patients without a supervising pharmacist or physician is prohibited.
extra_gated_fields:
  "I acknowledge this is NOT a doctor/pharmacist and NOT medical advice / ยอมรับว่าไม่ใช่แพทย์หรือเภสัชกรและไม่ใช่คำแนะนำทางการแพทย์": checkbox
  "I acknowledge the knowledge base is not yet pharmacist-reviewed (in review) / ยอมรับว่าข้อมูลยังไม่ผ่านการตรวจสอบโดยเภสัชกร (อยู่ระหว่างตรวจสอบ)": checkbox
  "I will have the information verified by a qualified professional before use / จะให้ผู้เชี่ยวชาญตรวจสอบข้อมูลก่อนนำไปใช้": checkbox
  "I accept the risk and the HAI-DEF Terms": checkbox
  "Organization / User": text
---

# pai1-medgemma-27b — Thai Pharmacy Assistant (v26-DPO, premium tier)

> ⚠️ **DRAFT — pharmacist & legal review pending.** The drug knowledge base behind this model is
> partly AI-generated and **not yet fully reviewed by a licensed pharmacist** (534 entries flagged).
> Released under an **acknowledge-before-download** gate. **Every output must be verified by a
> licensed pharmacist before real-world use.**

## What this model is

`pai1-medgemma-27b` is a **Thai community-pharmacy assistant** fine-tuned from
[`google/medgemma-27b-it`](https://huggingface.co/google/medgemma-27b-it) by **I C Develop Co., Ltd.**

**Its job:** help a pharmacy counter handle over-the-counter (OTC) requests in Thai — take a short
history, rule out drugs that are unsafe for that customer, recommend a suitable OTC option with dose
and cautions, and **refer to a doctor when red-flag symptoms appear**.

**Design principle — reason like a clinician internally, present as a pharmacy assistant externally.**
The model is trained to do the clinical reasoning a good pharmacist does (contraindication checks,
interaction awareness, red-flag triage) while **never claiming to be a physician and never issuing a
diagnosis** — presenting as a doctor is legally prohibited in Thailand.

**Tier:** this is the **premium tier** (higher accuracy). The smaller, faster, cheaper tier is
[`icdevelop/pai1-medgemma-4b`](https://huggingface.co/icdevelop/pai1-medgemma-4b). A smaller **FP8** build (~28 GB, near-identical quality) is at [`icdevelop/pai1-medgemma-27b-fp8`](https://huggingface.co/icdevelop/pai1-medgemma-27b-fp8).

**Multimodal:** inherits MedGemma's image input — intended for **medicine boxes, labels and
prescriptions**, *not* X-rays or lab films.

## Benchmarks

Version lineage (higher is better; every score below was produced by running **all** cases):

| Task | v20 | v24 | **v26-DPO (this release)** |
|---|---|---|---|
| lookup (no RAG) | 46.5 % | 43.0 % | **50.5 %** |
| lookup + RAG | 64.5 % | 65.3 % | **64.5 %** |
| deliberation | 29.2 % | 37.7 % | **38.3 %** |

**Reading the numbers:** v26-DPO is the strongest on drug-fact recall (**lookup 50.5 %**, +7.5 over
v24) while holding RAG performance and slightly improving clinical deliberation. Deliberation has
risen across the whole line (29.2 → 38.3) thanks to the deliberation training described below.

### How the benchmark was run (methodology)

All benchmarks are **deterministic** — the model is served on **vLLM** with `--seed 0 --enforce-eager`,
and answers are graded by an LLM judge (0–10). A case counts as **pass** only when the judge passes it
**and** no safety violation (`must_not`) is triggered. All benchmark cases are **held out** and
**decontaminated** from the training corpus, so scores are not inflated by memorisation.

| Benchmark | Cases | What it measures |
|---|---|---|
| `thai_pharma_bench_v2` (lookup) | 200 | Brand / route / interaction facts with **no retrieval**. Ground truth from the Thai TMT registry + KB. Pure drug-fact recall. |
| `thai_pharma_bench_v2_rag` (lookup + RAG) | 200 | Same questions, but real KB facts are retrieved and injected into the prompt — **the intended production setting**. |
| `deliberation_bench_v1` | 120 | The obvious/top retrieved drug is **contraindicated** for the customer's disclosed profile. The judge checks whether the model **rules it out, explains why, and recommends a safe alternative** — reasoning, not recitation. |

Before release, the merged model also passes a **6-prompt generation check** (headache+fever,
greeting, diarrhoea, allergic rhinitis, drug comparison, abdominal pain) verifying coherent Thai
output with no repetition or degeneration.

## How it was trained

- **Base model:** [`google/medgemma-27b-it`](https://huggingface.co/google/medgemma-27b-it) (Google MedGemma, Gemma-3 multimodal).
- **Stage 1 — LoRA SFT** (rank 32, 2 epochs, bf16, LR 3e-6) on `thai_pharma_v24` (~33 k dialogues).
  The **vision tower and multi-modal projector are frozen** — this is a text-domain adaptation, so
  image understanding stays exactly as MedGemma trained it.
- **Stage 2 — combined DPO** (β 0.1, `ld_alpha` 0.5 length-desensitisation, LR 1e-6, 100 steps) on
  6,648 preference pairs = 2,872 OTC-accuracy pairs + 3,776 **deliberation** pairs.
- **Training corpus** (`thai_pharma_v24`): curated Thai pharmacy dialogues — brand / route /
  interaction lookup, supplement & vitamin guidance, professional-register Q&A, TMT product grounding,
  drug-interaction cases, plus the deliberation set below. Exact-dedup + MinHash-LSH near-dedup, then
  **decontaminated against every evaluation set**.
- **Deliberation data** (`deliberate_v2`) — the key ingredient: multi-turn dialogues in which at least
  one retrieved candidate is *contraindicated* for the customer's disclosed profile (pregnancy,
  warfarin, gastric ulcer, driving, …). The target answer must take history → **rule out the unsafe
  option with a reason** → recommend a safe alternative → give dose/cautions from the retrieved facts
  → refer on red flags, and **ask for more history when information is insufficient** instead of
  guessing. A verifier rejects any turn that hard-diagnoses, impersonates a physician, or merely
  recites the record.
- **Deliberation preference pairs** (DPO): *chosen* = the clinician-style deliberation, *rejected* = a
  fluent but unsafe answer that recommends the contraindicated drug. The pairs are **length-balanced**
  (chosen/rejected token ratio ≈ 1.0) so the preference signal is **behaviour, not verbosity**.

## How to use

### Recommended: serve with vLLM and ground every answer with RAG

```bash
vllm serve icdevelop/pai1-medgemma-27b \
  --dtype bfloat16 --max-model-len 8192 \
  --served-model-name pai --seed 0 --enforce-eager

# the knowledge base used for grounding
huggingface-cli download icdevelop/thai-pharma-kb --repo-type dataset
```

### Getting the best results

1. **Always ground with RAG.** Retrieve the real drug facts (from the TMT registry / KB) for the
   drugs in question and inject them into the prompt. Lookup+RAG (64.5 %) is far above no-RAG
   behaviour, and grounded answers are what the model was trained to produce.
2. **Write the reference block and the question in Thai.** The model is purpose-built for Thai
   community pharmacy; Thai prompts get noticeably better answers.
3. **Let it take history.** If the customer's message lacks key details, the model is trained to ask
   follow-up questions — allow a multi-turn conversation rather than forcing a single-shot answer.
4. **Give it more than one candidate drug.** The deliberation training shows its value when the model
   must choose between options and exclude the unsafe one.
5. **Keep a pharmacist in the loop.** Treat every answer as a draft recommendation for a licensed
   pharmacist to confirm.

### Not recommended

- Using it **without retrieval** for factual drug questions (recall is much weaker).
- Asking for a **diagnosis**, or presenting it to end users as a doctor — it is trained to refuse and
  refer, and doing so is legally prohibited in Thailand.
- Radiology or laboratory image interpretation — out of scope.

## Data provenance & review status

| Source | Count | Status |
|---|---|---|
| TMT drug registry (Thai MoPH) | ~31,000 products | Official government data |
| Ingredient knowledge base | 1,918 ingredients | **AI-generated — not yet pharmacist-reviewed** (534 flagged for review) |

### Knowledge cutoff (data freshness)

The knowledge available to this model is bounded by the **RAG knowledge base**, not by the weights.

- **Knowledge base:** `icdevelop/thai-pharma-kb` · **cutoff: July 2026**
- Drugs, brands, supplements or regulatory changes registered after that date are **not** reflected.
  To advance the cutoff, re-pull the TMT registry, rebuild the KB, and update the dataset.

## Model format

**BF16** — full-precision merged weights, ~52 GB on disk.

## Limitations & safety

- **Not a medical device.** Not for diagnosis or treatment. Every output must be reviewed by a
  licensed pharmacist before it reaches a patient.
- **The knowledge base is not yet pharmacist-reviewed** and may contain inaccuracies (534 entries
  flagged). This is why the repository is gated behind an explicit acknowledgement.
- **Thai community-pharmacy domain only.** Image input is limited to medicine boxes, labels and
  prescriptions.
- The model can still make mistakes on rare drugs, complex poly-pharmacy and unusual interactions —
  RAG grounding reduces but does not eliminate this.

## License

A Model Derivative of [`google/medgemma-27b-it`](https://huggingface.co/google/medgemma-27b-it) under the
**Health AI Developer Foundations (HAI-DEF) Terms**. Redistribution requires attaching the Terms, the
§3.2 use restrictions, the NOTICE, and a modification notice.

**Not a medical device; not for clinical use without a licensed pharmacist.**
Contact: I C Develop Co., Ltd.