medgemma-anatomy-v1.2

MedGemma-4B Anatomy v1.2

Fine-tuned version of google/medgemma-4b-it specialized in anatomical and clinical reasoning.

Model Description

This model was fine-tuned on 183 high-quality medical Q&A pairs focusing on:

• Clinical anatomy
• Orthopedic injuries
• Neurological assessments
• Diagnostic approaches
• Management protocols

Training Details

• Base Model: google/medgemma-4b-it
• Training Data: 164 training samples, 19 validation samples
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Hardware: Google Colab T4 GPU
• Training Time: ~1.5-2 hours
• Parameters: 4B (32.7M trainable via LoRA)

Training Configuration

- Learning Rate: 1e-4
- Epochs: 5
- Batch Size: 1 (effective 16 with gradient accumulation)
- Max Sequence Length: 1024 tokens
- LoRA Rank: 16
- LoRA Alpha: 32

Usage

Direct Loading

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "krishna195/medgemma-anatomy-v1.2"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Ask a medical question
question = "A patient presents with shoulder pain after a fall. What is your diagnostic approach?"

prompt = f'''<start_of_turn>user
{question}<end_of_turn>
<start_of_turn>model
'''

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True,
    top_p=0.9
)

answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(answer.split("<start_of_turn>model")[-1])

4-bit Quantized Loading (Lower Memory)

from transformers import BitsAndBytesConfig

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=bnb_config,
    device_map="auto"
)

Performance

• Average Response Length: ~400 words
• Inference Speed: ~30-40 tokens/sec (T4 GPU)
• Memory Usage: ~6-8GB (full precision), ~3-4GB (4-bit)

Limitations

• Specialized in anatomy and orthopedics
• Not a replacement for professional medical advice
• Best for educational and reference purposes
• May be verbose on certain topics

Intended Use

• Medical education and training
• Clinical reasoning practice
• Anatomical reference assistant
• Medical exam preparation

NOT intended for: Clinical decision-making, patient diagnosis, or treatment planning without human oversight.

Training Data

Dataset consists of Q&A pairs extracted from medical textbooks with focus on:

• Fracture management
• Nerve injuries
• Joint anatomy
• Clinical presentations
• Diagnostic workflows

Average answer structure:

1. Overview & Pathophysiology
2. Clinical Presentation
3. Diagnostic Approach
4. Management Principles
5. Clinical Vignette

Citation

@misc{medgemma-anatomy-v1.2,
  title={MedGemma-4B Anatomy Fine-tuned Model},
  author={Your Name},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{krishna195/medgemma-anatomy-v1.2}}
}

License

Apache 2.0 (following base model license)

Contact

For questions or issues, please open an issue on the model repository.