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-4b-it
pipeline_tag: image-text-to-text
library_name: transformers
language: [th, en]
tags: [medical, pharmacy, thai, medgemma, gemma3, lora, rag, dpo]
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** และผลลัพธ์ทุกอย่างต้องผ่านการตรวจสอบโดยเภสัชกรที่มีใบอนุญาต 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-4b — Thai Pharmacy Assistant (v25-DPO, standard tier)

> ⚠️ **DRAFT — pharmacist & legal review pending. The drug knowledge base is AI-generated and not yet
> pharmacist-reviewed (review in progress).** Released under an **acknowledge-before-download** gate.
> All output must be verified by a licensed pharmacist before real use.

A Thai community-pharmacy assistant (4B) fine-tuned from `google/medgemma-4b-it` by **I C Develop Co., Ltd.** — the
small, fast, economical tier (premium tier: `icdevelop/pai1-medgemma-27b`).

# ⚠️ This model MUST be used with RAG + the Knowledge Base
It is **not a self-contained drug store.** Queried alone it will **generate inaccurate drug info.** For
every query, retrieve real facts from `icdevelop/thai-pharma-kb` and inject them into the prompt. See
`rag_example.py`. The 4B base is too small to read context on its own — **use this fine-tuned model, not the base.**

## Benchmarks (deterministic, seed 0; verified case counts 200/200/120)

| Usage | lookup | lookup + RAG | deliberation |
|---|---|---|---|
| **v25-DPO (this release)** alone (no RAG) | 37.5% | — | — |
| **v25-DPO + RAG (intended use)** | — | **64.0%** | **48.3%** |
| v24-DPO + RAG | — | 62.5% | 40.0% |
| v24 SFT + RAG | — | 53.0% | 44.2% |
| v20 + RAG (previous) | — | 55.0% | 37.5% |

**What's new in v25-DPO:** SFT on the improved deliberation set (`deliberate_v2`) **followed by a DPO
(Direct Preference Optimization) refinement stage**. This release **beats v20 on all three axes** and has
the **best 4B lookup+RAG score yet (55% → 62.5%)** — RAG-grounded is the production setting. Deliberation
lineage: v13 22.5% → v20 37.5% → v24-SFT 44.2% → **v25-DPO 40.0%** (DPO traded ~4 pt of pure deliberation
to recover lookup/RAG well beyond v20). Use RAG always; deliberation stays clinician-style (rules out a
contraindicated drug, recommends a safe alternative, asks for history when info is insufficient).

## How it was trained (for reproducibility & trust)

- **Base:** `google/medgemma-4b-it` (Google MedGemma, Gemma-3 multimodal). Fine-tuned by **I C Develop Co., Ltd.**
- **Method:** LoRA (rank 64) via LLaMA-Factory. **Vision tower + multi-modal projector frozen** (text-only
  adaptation). **Two stages: SFT** (`thai_pharma_v24`, 33,760 examples, 2 epochs, LR 1e-4) **then a combined DPO**
  (preference set `dpo_v25` = 2,872 OTC-accuracy pairs + 3,776 deliberation pairs, sigmoid loss, β 0.1,
  LR 5e-6). The deliberation pairs (chosen = deliberates, rejected = recites the contraindicated drug)
  reward reasoning directly, so deliberation rises instead of drifting down.
- **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, and the **deliberation set** below. Exact + MinHash-LSH dedup, then
  **decontaminated** against every eval set (train/test leakage removed → honest scores).
- **Deliberation data (`deliberate_v2`):** multi-turn dialogues where ≥1 retrieved candidate is
  *contraindicated* for the customer's disclosed profile. The teacher takes history → rules out the
  unsafe option **with a reason** → recommends a safe alternative → gives dose/cautions from facts →
  refers on red flags, and **asks for more history when info is insufficient**. A verifier rejects any
  turn that hard-diagnoses, impersonates a physician, or merely recites — enforcing the design goal:
  **reason like a clinician internally, present as a pharmacy assistant externally** (legally required in Thailand).

## How the benchmark was run (why these scores mean something)

Deterministic — served on **vLLM** with `--seed 0 --enforce-eager`, scored by an LLM judge (0-10; "pass"
= judge ≥ threshold), held-out and decontaminated from training.
- **`thai_pharma_bench_v2` (lookup, 200):** brand/route/interaction facts; ground truth = TMT + KB.
- **`thai_pharma_bench_v2_rag` (lookup+RAG, 200):** same, with real KB facts retrieved into the prompt — the production setting.
- **`deliberation_bench_v1` (120):** the top retrieved drug is *contraindicated* for the patient; judge
  scores whether the model rules it out and recommends a safe alternative (reasoning, not recitation).

## How to use (RAG — see `rag_example.py`)

```bash
vllm serve icdevelop/pai1-medgemma-4b --dtype bfloat16 --max-model-len 8192 \
    --served-model-name pai --seed 0 --enforce-eager
huggingface-cli download icdevelop/thai-pharma-kb --repo-type dataset
```

On **every** call, retrieve real drug facts from the TMT registry / KB and inject them into the prompt
(Thai reference block + Thai question). For maximum benefit: always ground with RAG, keep the retrieved
facts in the prompt, and let the model take history before recommending. See `rag_example.py`.

## Data provenance & review status

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

## RAG knowledge cutoff (data freshness)

Knowledge is bounded by the RAG knowledge base, **not** the weights.
- **Knowledge base:** `icdevelop/thai-pharma-kb` · **knowledge cutoff: July 2026**
- Drugs/brands/regulatory changes after that date are not reflected.

## Limitations & safety

- **Must always be used with RAG.** The model alone recalls drug info poorly.
- **Not a medical device.** Every output must be reviewed by a licensed pharmacist. The KB is not yet
  pharmacist-reviewed and may be inaccurate (534 flagged). Thai community-pharmacy domain only; image
  input = medicine boxes / labels / prescriptions — not X-rays or lab films.

## License (HAI-DEF)

A Model Derivative of `google/medgemma-4b-it` under the Health AI Developer Foundations Terms. Redistribution requires
attaching the Terms + the §3.2 use restrictions + the NOTICE + a modification notice.

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