Model Card for Vietnamese Medical NER (Qwen2.5-7B SFT + GRPO)

This model is a generative Large Language Model (based on the Qwen/Qwen2.5-7B architecture) strictly fine-tuned for Vietnamese Medical Named Entity Recognition (NER). It accurately extracts Symptoms, Diseases, Medical Procedures, Drugs, etc. from Vietnamese medical texts and outputs them in a highly structured JSON format.

Model Details

Model Description

This model transitions away from traditional token-classification BERT models by utilizing an LLM as an intelligent extraction agent. It reads Vietnamese medical texts (such as clinical notes, user questions, or news) and outputs structured JSON lists.

The model was trained using an advanced two-stage alignment pipeline:

1. Stage 1: Supervised Fine-Tuning (SFT) to teach the model the strict JSON extraction schema and Vietnamese medical terminology.
2. Stage 2: Group Relative Policy Optimization (GRPO) using Reinforcement Learning (RL) with LoRA. The model was mathematically penalized via shifted-bound and exponential scaling reward functions for hallucinations and missed entities, forcing it to achieve near-perfect Precision, Recall, and Type Accuracy.

• Developed by: PeterPaker123
• Model type: Causal Language Model (Generative NER)
• Language(s) (NLP): Vietnamese (vi)
• License: Apache 2.0 (Derived from Qwen base models)
• Finetuned from model: Qwen/Qwen2.5-7B-Instruct

Uses

Direct Use

Unlike traditional NER models constrained to a fixed set of predefined classes, this model supports dynamic, prompt-defined entity extraction. It acts as an intelligent agent that extracts any medical or clinical entity type specified in the system prompt.

Commonly targeted categories include (but are not limited to):

• SYMPTOM_AND_DISEASE: Symptoms, signs of illness, diseases, or medical conditions.
• MEDICAL_PROCEDURE: Medical procedures, surgeries, therapies, or diagnostic methods.
• DRUG: Names of medicines, drugs, vitamins, or supplements.
• BODY_PART / ANATOMY: Anatomical locations, organs, or body structures.
• DOSAGE & MEASUREMENT: Medical quantities, medication dosages, or frequencies.
• DEMOGRAPHICS: Patient-specific identifiers (e.g., Age, Gender, Occupation).

How it works: You simply define the desired categories and their descriptions in the system prompt, and the model will strictly adhere to extracting only those requested types into the required JSON schema.

Downstream Use

This model can be integrated into downstream healthcare applications such as:

• Medical chatbot prompt enrichment (RAG pipelines).
• Automated summarization of telehealth transcripts.
• Clinical knowledge graph construction.

Out-of-Scope Use

• Medical Diagnosis or Triage: This model is strictly an extraction tool. It does not diagnose, prescribe, or provide medical advice.
• General Conversation: The model has been aggressively aligned to output JSON lists. It will perform poorly at standard conversational AI tasks.
• Other Entity Types: The model was trained to ignore Demographic data (Names, Ages, Locations, Occupations, Dates). Attempting to extract these will likely yield empty lists.

Bias, Risks, and Limitations

• Clinical Disclaimer: This model is an AI research tool and must not be used for automated medical diagnosis, treatment planning, or life-or-death triage without human oversight.
• Language Limitation: The model is exclusively trained for the Vietnamese language.
• Hallucination Residuals: While GRPO significantly reduces hallucinations, Generative LLMs may occasionally hallucinate or alter the spelling of extracted text.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Extracted entities in a clinical setting must be validated by a human medical professional or a deterministic cross-checking system.

How to Get Started with the Model

Use the code below to get started with the model.

import json
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "PeterPaker123/Qwen2.5-7B-Vietnamese-Medical-NER-GRPO"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="auto")

prompt = [
  {
    "role": "system",
    "content": "You are a medical expert. Your task is to identify and extract medical entities from the given text.\nThe entity types to extract are:\n- SYMPTOM_AND_DISEASE: Symptoms, signs of illness, diseases, or medical conditions.\n- MEDICAL_PROCEDURE: Medical procedures, surgeries, therapies, or diagnostic methods.\n- DRUG: Names of medicines, drugs, vitamins, or supplements.\n\nReturn the result as a JSON list containing dictionaries with \"entity\" and \"type\" keys. If no relevant entities are found, return an empty list []"
  },
  {
    "role": "user",
    "content": "Text: Trẻ sơ sinh 2 tháng tuổi vàng mặt , hay khóc thét có phải bị vàng da nhân não không ?"
  }
]

inputs = tokenizer.apply_chat_template(prompt, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.0)

response = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
print(response)
# Expected Output: [{"entity": "vàng mặt", "type": "SYMPTOM_AND_DISEASE"}, {"entity": "khóc thét", "type": "SYMPTOM_AND_DISEASE"}, {"entity": "vàng da nhân não", "type": "SYMPTOM_AND_DISEASE"}]

Training Details

Training Data

This model was trained on a consolidated corpus derived from four premier Vietnamese medical datasets. The original datasets contained diverse entity tags which were mapped down to our three primary clinical targets (SYMPTOM_AND_DISEASE, MEDICAL_PROCEDURE, DRUG).

Data Split: 75% used for Phase 1 (SFT) and 25% used for Phase 2 (GRPO/RL).

Training Procedure

Preprocessing [optional]

• Prompts were formatted into strict ChatML format using text-only templates.
• Target entities outside the three core classes were stripped from the ground truth to force the model to ignore non-medical demographic data.

Training Hyperparameters

• Training regime: bf16 mixed precision
• Phase 1 (SFT): Full parameter fine-tuning, Max Seq Len 4096, Batch Size 2, Gradient Accumulation 8, LR 2e-5, Cosine Scheduler.
• Phase 2 (GRPO): LoRA target modules (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj), Rank 16, Alpha 32. 4 Generations per prompt, Temperature 0.5. Full AdamW optimizer.

Model Examination

Technical Specifications

Model Architecture and Objective

Qwen2.5-7B Architecture. Causal Language Modeling objective heavily modified via RLHF/GRPO to act as an entity extraction classifier.

Hardware

NVIDIA RTX PRO 6000 Blackwell

Software

• transformers
• trl (SFTTrainer, GRPOTrainer)
• peft
• torch (with sdpa attention)

Citation

BibTeX:

@inproceedings{vietbioner,
    title = "{A Named Entity Recognition Corpus for Vietnamese Biomedical Texts to Support Tuberculosis Treatment}",
    author = "Phan, Uyen and Nguyen, Phuong and Nguyen, Nhung",
    booktitle = "Proceedings of the 13th Language Resources and Evaluation Conference",
    year = "2022",
    publisher = "European Language Resources Association"
}

@inproceedings{PhoNER_COVID19,
    title = {{COVID-19 Named Entity Recognition for Vietnamese}},
    author = {Thinh Hung Truong and Mai Hoang Dao and Dat Quoc Nguyen},
    booktitle = {Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies},
    year = {2021}
}

@article{vimq2023,
  title={ViMQ: A Vietnamese Medical Question Dataset for Healthcare Dialogue System Development},
  author={Vu, Hieu M. and Phan, Long and Nguyen, Nhung and others},
  year={2023}
}

@article{10.4108/eetinis.v11i3.5221,
    author={Pham Van Duong and Tien-Dat Trinh and Minh-Tien Nguyen and Huy-The Vu and Minh Chuan Pham and Tran Manh Tuan and Le Hoang Son},
    title={ViMedNER: A Medical Named Entity Recognition Dataset for Vietnamese},
    journal={EAI Endorsed Transactions on Industrial Networks and Intelligent Systems},
    volume={11},
    number={4},
    publisher={EAI},
    year={2024},
    doi={10.4108/eetinis.v11i3.5221}
}

Glossary

• NER: Named Entity Recognition
• SFT: Supervised Fine-Tuning
• GRPO: Group Relative Policy Optimization (An RL algorithm optimized for Generative Models)
• LoRA: Low-Rank Adaptation