dousery/medical-reasoning-gpt-oss-20b

Medical Reasoning GPT-OSS-20B

Model Description

This is a fine-tuned version of unsloth/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
• Safety-Focused: Trained with responsible AI practices for healthcare applications
• Large Scale: 20B parameters for comprehensive medical knowledge
• Ready-to-Use: Full model (not just LoRA adapter) - no additional setup required

🚀 Quick Start

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True
)

model_id = "dousery/medical-reasoning-gpt-oss-20b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    quantization_config=quantization_config,
    device_map="auto"
)

prompt = "A patient has symptoms of fever and cough. What could be the diagnosis?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.7,
    do_sample=True,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.eos_token_id
)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(generated_text)

📊 Training Details

Training Data

• Dataset: 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 (Low-Rank Adaptation) fine-tuning
• LoRA Rank: 8
• Learning Rate: 2e-4
• Batch Size: 4 (effective)
• Epochs: 1
• Hardware: NVIDIA B200 (4x GPUs)
• Framework: Unsloth + TRL
• Final Training Loss: 0.88

Model Architecture

• Parameters: 20.9 billion
• Architecture: GPT-OSS (Transformer-based)
• Context Length: 1,024 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.

📈 Performance Metrics

• Training Loss: 10.78 → 0.88 (significant improvement)
• Convergence: Stable training with consistent loss reduction
• Reasoning Quality: Maintains logical chain-of-thought structure

🛠️ Technical Requirements

Minimum Requirements

• GPU Memory: 16GB+ VRAM recommended
• RAM: 32GB+ system memory
• Storage: 40GB+ free space

Optimized for

• Inference: FP16/BF16 precision
• Frameworks: Transformers, Unsloth, TRL
• Hardware: NVIDIA GPUs with Compute Capability 7.0+

📜 License

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

🙏 Acknowledgments

• Base Model: unsloth/gpt-oss-20b
• Training Framework: Unsloth
• Dataset: Freedomintelligence
• Infrastructure: Modal Labs for GPU compute

📞 Contact

For questions, issues, or collaboration opportunities, please reach out through the HuggingFace community discussions or my Linkedin account : Linkedin

Version: 1.0
Release Date: January 2025
Model Type: Causal Language Model
Training Infrastructure: Modal Labs B200 GPU Cluster