You can deploy this model and start earning money today!

serhanayberkkilic/
qwen3-14b-physiotherapy-lora

Name: qwen3-14b-physiotherapy-lora
Availability: InStock

Views

No views yet

Qwen3-14B Physiotherapy Evidence-Based QA (LoRA)

🏥 A bilingual (English/Turkish) Large Language Model fine-tuned for evidence-based physiotherapy and clinical rehabilitation Q&A

📋 Model Overview

This model is a LoRA (Low-Rank Adaptation) fine-tuned version of Qwen3-14B, specifically trained on evidence-based physiotherapy literature. It provides accurate, clinically-relevant answers for physiotherapy and rehabilitation questions in both English and Turkish.

Key Features

Feature	Description
🌍 Bilingual	Supports both English and Turkish
📚 Evidence-Based	Trained on peer-reviewed clinical literature
⚡ Efficient	LoRA adapter (~500MB) - minimal VRAM with 4-bit quantization
🔬 Clinical Focus	Specialized for physiotherapy and rehabilitation
💬 Conversational	Optimized for Q&A format

🎯 Intended Use Cases

Clinical Decision Support: Evidence-based treatment recommendations
Patient Education: Generating educational content about conditions and treatments
Research Assistance: Literature review and knowledge synthesis
Medical Training: Educational tool for physiotherapy students
Healthcare Applications: Integration into clinical systems and chatbots

📊 Model Specifications

Specification	Value
Base Model	`unsloth/Qwen3-14B`
Fine-tuning Method	LoRA (Low-Rank Adaptation)
LoRA Rank (r)	32
LoRA Alpha	32
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Trainable Parameters	~1-2% of total
Quantization	4-bit (BitsAndBytes NF4)
Max Sequence Length	2048 tokens
Training Framework	Unsloth + TRL + PEFT

📁 Training Data

This model was trained on the Physiotherapy Evidence QA Dataset:

Statistic	Value
Total Q&A Pairs	143,711
Training Conversations	~287,422 (EN + TR)
Languages	English & Turkish
Domain	Evidence-based Physiotherapy
Sources	Peer-reviewed clinical literature

Dataset Topics Include:

Musculoskeletal conditions (back pain, joint disorders, etc.)
Neurological rehabilitation
Sports injuries and rehabilitation
Post-surgical rehabilitation
Outcome measures and assessment tools (PRTEE, ODI, etc.)
Therapeutic exercises and interventions
Manual therapy techniques

🚀 Quick Start

Installation

pip install transformers peft accelerate bitsandbytes torch

Inference Code

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

# Configuration
base_model_name = "unsloth/qwen3-14b-unsloth-bnb-4bit"
lora_adapter = "serhanayberkkilic/qwen3-14b-physiotherapy-lora"

# 4-bit quantization
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(lora_adapter, trust_remote_code=True)

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, lora_adapter)
model.eval()

# Generate response
def generate(question: str, max_tokens: int = 512):
    messages = [{"role": "user", "content": question}]
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(text, return_tensors="pt").to(model.device)
    
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_tokens,
            temperature=0.7,
            top_p=0.9,
            do_sample=True,
        )
    
    response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
    return response

# Example usage
question = "What are the evidence-based treatments for chronic low back pain?"
answer = generate(question)
print(answer)

💡 Example Outputs

English Example

Question: What is the PRTEE questionnaire and how is it used?

Answer: The Patient-Rated Tennis Elbow Evaluation (PRTEE) is a validated outcome measure designed to assess pain and functional disability in patients with lateral epicondylitis. It consists of 15 items divided into two subscales: pain (5 items) and function (10 items). The function subscale is further divided into specific and usual activities. Scores range from 0-100, with higher scores indicating greater pain and disability. The PRTEE has demonstrated excellent reliability (ICC > 0.90) and validity for measuring treatment outcomes in lateral epicondylitis...

Turkish Example

Soru: Bel ağrısı için kanıta dayalı fizyoterapi tedavi önerileri nelerdir?

Yanıt: Bel ağrısı tedavisinde kanıta dayalı fizyoterapi yaklaşımları arasında egzersiz terapisi, manuel terapi, hasta eğitimi ve davranışsal yaklaşımlar yer almaktadır. Egzersiz terapisi en güçlü kanıt düzeyine sahip olup, core stabilizasyon egzersizleri, genel kondisyon egzersizleri ve spesifik gövde güçlendirme programları önerilmektedir...

⚙️ Hardware Requirements

Configuration	VRAM Required	Recommended GPU
4-bit (Default)	~10-12 GB	RTX 3090, RTX 4090, A100
8-bit	~18-22 GB	A100, H100
16-bit (FP16)	~28-32 GB	A100 80GB, H100

⚠️ Limitations & Disclaimers

Limitations

Not a replacement for clinical judgment: This model should be used as a supportive tool only
Training data cutoff: Knowledge is limited to the training dataset
May generate inaccurate information: Always verify with primary sources
Language quality: Turkish responses may occasionally include English terms

Ethical Considerations

This model is intended for educational and research purposes
Should not be used for direct patient diagnosis or treatment without professional oversight
Users should be aware of potential biases in medical literature

📈 Training Details

Training Configuration

Parameter	Value
Optimizer	AdamW 8-bit
Learning Rate	2e-4
LR Scheduler	Cosine
Batch Size	Per device varies by GPU
Gradient Accumulation	Adjusted for effective batch size
Epochs	1-2 passes over data
Gradient Checkpointing	Unsloth optimized

Training Infrastructure

Hardware: NVIDIA GPU (RTX 4090 / A100)
Framework: Unsloth + HuggingFace TRL + PEFT
Precision: 4-bit quantization (NF4)

📝 Citation

If you use this model in your research, please cite:

bibtex

@misc{kilic2024physiolora,
  author = {Kilic, Serhan Ayberk},
  title = {Qwen3-14B Physiotherapy Evidence-Based QA LoRA},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/serhanayberkkilic/qwen3-14b-physiotherapy-lora}
}

👤 Author

Serhan Ayberk Kılıç

🤗 HuggingFace: @serhanayberkkilic
📧 Contact: For research collaborations and inquiries

📄 License

This model is released under the Apache 2.0 License. See LICENSE for more details.

🙏 Acknowledgments

Qwen Team for the excellent Qwen3 base model
Unsloth for optimized training framework
HuggingFace for the transformers ecosystem
The physiotherapy research community for evidence-based literature

Framework Versions

PEFT: 0.18.1
Transformers: 4.45.0+
PyTorch: 2.0+
Unsloth: Latest