seiya/Phi-4-mini-instruct-qlora-adapter-jp-disease-finding

Phi-4-mini-instruct QLoRA Adapter (jp-disease-finding)

TL;DR – A 4-bit QLoRA adapter that specializes the small Phi-4-mini-instruct model for span-level extraction of disease names and clinical findings from Japanese medical journal text.
The adapter was trained on a subset (vol. 98–102) of the jp-disease-finding-dataset.

Access & Usage Conditions

This repository is a manual-approval gated model. You may browse the file list, but downloading any file requires an approved access request Hugging Face. Requests are reviewed manually by the authors; please allow 1–3 business days for a decision.

1. Model Details

🚀 Notebook / Code Quick-start
GitHub repo

2. Intended Use & Scope

• Direct use: automatic extraction of Disease and Finding spans from Japanese medical literature for database construction or downstream NLP pipelines.
• Users: clinical NLP researchers, medical informatics engineers, healthcare data scientists.
• Out-of-scope: deployment without domain-expert supervision in clinical decision support systems.

3. How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE    = "microsoft/Phi-4-mini-instruct"
ADAPTER = "seiya/Phi-4-mini-instruct-qlora-adapter-jp-disease-finding"

model = AutoModelForCausalLM.from_pretrained(
    BASE, torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(model, ADAPTER)

tokenizer = AutoTokenizer.from_pretrained(BASE)
prompt = "慢性関節リウマチの診断と管理……"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(out[0], skip_special_tokens=True))

For an end-to-end demonstration (prompt design, post-processing, evaluation), see the notebook linked above.

4. Limitations & Bias

• Trained on journal language; may underperform on electronic health records or layperson text.
• Japanese only – no guarantee of accuracy on other languages.
• Hallucination and boundary errors are possible; always validate critical outputs.

---

Acknowledgment

This work was supported by JSPS KAKENHI Grant Number JP22K12263.

Citation

If you use this dataset, please cite:
Currently under preparation...

---

This model was trained 2x faster with Unsloth and Huggingface's TRL library.