---
-language:
- en
license: apache-2.0
base_model: openai/gpt-oss-20b
tags:
- medical
- reasoning
- healthcare
- fine-tuned
- gpt-oss
- clinical
- diagnosis
- generated_from_trainer
- sft
- trl
datasets:
- Freedomintelligence/medical-o1-reasoning-SFT
pipeline_tag: text-generation
library_name: transformers
widget:
- text: "A patient presents with fever of 39°C, severe headache, and neck stiffness. What could be the possible diagnosis?"
  example_title: "Medical Diagnosis"
- text: "Explain the pathophysiology of myocardial infarction and its treatment options."
  example_title: "Medical Education"
- text: "A 45-year-old patient with chest pain and shortness of breath. Analyze the differential diagnosis."
  example_title: "Clinical Reasoning"
model-index:
- name: medical-reasoning-gpt-oss-20b
  results:
  - task:
      type: text-generation
      name: Medical Text Generation
    dataset:
      type: medical-reasoning
      name: Medical O1 Reasoning SFT
    metrics:
    - type: perplexity
      name: Training Loss
      value: 1.22
---

# Medical Reasoning GPT-OSS-20B

## Model Description

This is a fine-tuned version of [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) specifically optimized for medical reasoning and clinical decision-making. The model has been trained on high-quality medical reasoning datasets to provide accurate and thoughtful responses to medical queries.

## 🏥 Key Features

- **Medical Expertise**: Specialized in medical reasoning, diagnosis, and clinical decision-making
- **Complex Reasoning**: Uses chain-of-thought reasoning for medical problems
- **Adapter-Only Training**: Only LoRA layers are trained, base model remains frozen
- **Efficient**: Lightweight fine-tuning, smaller storage footprint
- **Ready-to-Use**: Requires base model + adapter for inference

## 🚀 Quick Start

```python
#pip install torch --index-url https://download.pytorch.org/whl/cu128
#pip install "trl>=0.20.0" "peft>=0.17.0" "transformers>=4.55.0"

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

base_model_name = "openai/gpt-oss-20b"
adapter_name = "dousery/medical-reasoning-gpt-oss-20b"

tokenizer = AutoTokenizer.from_pretrained(base_model_name)

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

model = PeftModel.from_pretrained(base_model, adapter_name)
model = model.merge_and_unload()

messages = [
    {"role": "system", "content": "You are a medical reasoning assistant."},
    {"role": "user", "content": (
        """A 55-year-old man has chest pain and elevated troponin I without ST elevation.
         What is the diagnosis and what additional test would you order next?"""
    )}
]

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

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

outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    temperature=0.2,
    do_sample=False
)

raw_output = tokenizer.decode(outputs[0], skip_special_tokens=False)

#  PARSING THE OUTPUT
thinking_pattern = r"<\|end\|><\|start\|>assistant<\|channel\|>analysis<\|message\|>(.*?)<\|end\|>"
final_pattern = r"<\|start\|>assistant<\|channel\|>final<\|message\|>(.*?)<\|return\|>"

thinking_match = re.search(thinking_pattern, raw_output, re.DOTALL)
final_match = re.search(final_pattern, raw_output, re.DOTALL)

thinking_text = thinking_match.group(1).strip() if thinking_match else "N/A"
final_text = final_match.group(1).strip() if final_match else "N/A"

print("Thinking:", thinking_text)
print("\nFinal:", final_text)
```

## 📊 Training Details

### Training Data
- **Dataset**: [Freedomintelligence/medical-o1-reasoning-SFT](https://huggingface.co/datasets/Freedomintelligence/medical-o1-reasoning-SFT)
- **Language**: English
- **Size**: 19,704 medical reasoning examples
- **Format**: Question-Answer pairs with complex chain-of-thought reasoning

### Training Configuration
- **Base Model**: unsloth/gpt-oss-20b (20B parameters)
- **Training Method**: LoRA (adapter-only fine-tuning)
- **LoRA Rank**: 8
- **Learning Rate**: 5e-5
- **Batch Size**: 4 per device, gradient_accumulation_steps=4
- **Epochs**: 2
- **Max Sequence Length**: 2048
- **LR Scheduler**: Cosine, warmup_ratio=0.05
- **Final Training Loss**: 1.22

### Model Architecture
- **Parameters**: 20.9 billion
- **Architecture**: GPT-OSS (Transformer-based)
- **Context Length**: 2.048 tokens
- **Trainable Parameters**: 3.98M (0.02% of total)

## 🎯 Intended Use

### Primary Use Cases
- **Medical Education**: Explaining medical concepts and procedures
- **Clinical Reasoning**: Analyzing symptoms and differential diagnosis
- **Research Support**: Assisting in medical research and literature review
- **Decision Support**: Providing reasoning for clinical decisions (with human oversight)


## ⚠️ Important Disclaimers

- **Not a Medical Device**: This model is for educational and research purposes only
- **Human Oversight Required**: All medical decisions should involve qualified healthcare professionals
- **Accuracy Not Guaranteed**: Model outputs should be verified against current medical literature
- **Regional Variations**: Training data may not reflect all regional medical practices

## 🔍 Evaluation

The model demonstrates strong performance in:
- Medical concept explanation
- Differential diagnosis reasoning
- Treatment option analysis
- Pathophysiology understanding

**Note**: Comprehensive clinical evaluation is ongoing. Always validate outputs with current medical guidelines.

## 🛠️ Technical Requirements

### Minimum Requirements
- **GPU Memory**: 16GB+ VRAM recommended
- **RAM**: 32GB+ system memory
- **Storage**: 40GB+ free space


## 📜 License

This model is released under the Apache 2.0 license. Please review the license terms before commercial use.

## 🙏 Acknowledgments

- **Base Model**: [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b)
- **Adapter/Training**: [dousery/medical-reasoning-gpt-oss-20b]()
- **Dataset**: [Freedomintelligence](https://huggingface.co/Freedomintelligence)
- **Infrastructure**: Modal Labs for GPU compute