pkhare/qwen3-8b-biomedical

Qwen3-8B for Biomedical Language Processing

This research purpose model is a fine-tuned version of Qwen/Qwen3-8B for biomedical NLP tasks, trained using instruction tuning with QLoRA. It has been optimized for Named Entity Recognition (NER), Relation Extraction (RE), Natural Language Inference (NLI), and Document Classification in the medical domain.

📚 GitHub Repository: BioMedical-LLM

This work reproduces and extends "Exploring the Effectiveness of Instruction Tuning in Biomedical Language Processing" by adapting the approach to Qwen3-8B (replacing the original Llama2), leveraging Qwen3's reasoning capabilities.

Model Description

• Architecture: Qwen3-8B (8.2B parameters, 36 layers, 32K context)
• Training Method: QLoRA (4-bit quantization + LoRA rank 16)
• Training Data: instruction-formatted biomedical examples (as in the original work, data reproduction approach detailed on the GitHub page of BioMedical-LLM)
• Hardware: NVIDIA A100 80GB GPU (RunPod.io)
• Training Time: 3 epochs (~20 hours)

Supported Tasks

• Named Entity Recognition: Diseases, chemicals, genes, and more
• Relation Extraction: Treatment-problem, test-problem relations (8 clinical relation types)
• Natural Language Inference: Medical statement entailment/contradiction
• Document Classification: Cancer hallmarks, genetic associations (multi-label)

Quick Start

Basic Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

Load base model

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-8B",
    device_map="auto",
    trust_remote_code=True
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    "Qwen/Qwen3-8B",
    trust_remote_code=True
)

# Load LoRA adapters
model = PeftModel.from_pretrained(
    base_model,
    "pkhare/qwen3-8b-biomedical"
)

# Example: Named Entity Recognition
prompt = "Extract disease entities from: Patient diagnosed with diabetes and hypertension."
messages = [
    {"role": "system", "content": "You are a helpful medical AI assistant."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(text, return_tensors="pt").to("cuda")
outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    do_sample=True,
    temperature=0.7,
    top_p=0.8,
    top_k=20
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Pipeline Usage

from transformers import pipeline

# Initialize pipeline
generator = pipeline(
    "text-generation",
    model="pkhare/qwen3-8b-biomedical",
    device="cuda"
)

# Named Entity Recognition
ner_prompt = "Extract disease entities from: Patient diagnosed with diabetes mellitus and hypertension."
output = generator(
    [{"role": "user", "content": ner_prompt}],
    max_new_tokens=128,
    do_sample=True,
    temperature=0.7,
    top_p=0.8,
    top_k=20,
    return_full_text=False
)[0]
print(output["generated_text"])

Usage Examples

Named Entity Recognition (NER)

prompt = """Extract disease entities from the following text:
Patient diagnosed with diabetes mellitus and hypertension.

Format: List each disease on a new line."""

messages = [
    {"role": "system", "content": "You are a helpful medical AI assistant."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")
outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    do_sample=True,
    temperature=0.7,
    top_p=0.8,
    top_k=20
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Relation Extraction (RE)

prompt = """Given the following sentence, identify the relation between the treatment and problem:
"The patient was treated with metformin for type 2 diabetes."

Possible relations: TrIP (Treatment improves Problem), TeRP (Test reveals Problem), 
PIP (Problem indicates Problem), TrAP (Treatment causes Problem), 
TrCP (Treatment conducted for Problem), TrWP (Treatment worsens Problem), 
TrNAP (Treatment not administered because of Problem), 
TeCP (Test conducted to investigate Problem), No Relations

Answer:"""

messages = [
    {"role": "system", "content": "You are a helpful medical AI assistant."},
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")
outputs = model.generate(
    **inputs,
    max_new_tokens=64,
    do_sample=True,
    temperature=0.7,
    top_p=0.8,
    top_k=20
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Ethical Considerations

• ⚠️ Not for clinical use: This model and work is only for research purposes only and should not be used for medical diagnosis or treatment decisions without expert oversight
• ⚠️ Not fit for commercial purpose. This is only an experimental and research purpose model.
• Privacy: Do not input patient-identifiable information (PHI/PII).
• Verification: All model outputs should be verified by qualified medical professionals.
• Regulatory: Not FDA approved or clinically validated by any organisation.

Codebase

The code for this work is available at BioMedical-LLM

This research work is based on original work done by

@article{rohanian2024exploring,
  title = {Exploring the Effectiveness of Instruction Tuning in Biomedical Language Processing},
  author = {Rohanian, Omid and Nouriborji, Mohammadmahdi and Kouchaki, Samaneh and Nooralahzadeh, Farhad and Clifton, Lei and Clifton, David A},
  journal = {Artificial Intelligence in Medicine},
  volume = {158},
  pages = {103007},
  year = {2024},
  publisher = {Elsevier},
  doi = {10.1016/j.artmed.2024.103007},
  url = {https://www.sciencedirect.com/science/article/pii/S0933365724002495},
  issn = {0933-3657}
}