arunimas1107/gpt-oss-medical

Model Card for Medical GPT-OSS-20B LoRA Adapter

Model Details

Model Description

This model is a LoRA adapter fine-tuned on openai/gpt-oss-20b using the PEFT library. It is optimized for medical domain tasks such as question answering, summarization, and knowledge retrieval in healthcare contexts. The adapter modifies the base model with efficient fine-tuning techniques while retaining the general-purpose reasoning capabilities of the underlying 20B parameter model.

• Developed by: Arunima Surendran
• Funded by: E2E Cloud
• Shared by: Arunima Surendran
• Model type: Large Language Model with LoRA adapter (20B base)
• Language(s): English
• License: Apache 2.0
• Finetuned from model: openai/gpt-oss-20b
• Finetuned using: NVIDIA 2xH200 for 12+ hours on E2E Cloud TIR Instance

Model Sources

• Repository: arunimakanavu/gpt-oss-medical
• Paper [optional]: [More Information Needed]
• Demo [optional]: [More Information Needed]

Requirements

torch>=2.0.0
transformers @ git+https://github.com/huggingface/transformers.git
datasets>=2.12.0
peft>=0.10.0
accelerate>=0.22.0
bitsandbytes>=0.41.0
sentencepiece>=0.1.99

Uses

Direct Use

• Medical Q&A
• Clinical text summarization
• Educational content generation in healthcare

Downstream Use

• Integrating into RAG pipelines with domain-specific medical knowledge bases
• Deployment in medical chatbots (for informational purposes only)

Out-of-Scope Use

• Direct clinical decision-making or diagnostic tools without human oversight
• High-stakes medical applications without proper validation

Bias, Risks, and Limitations

The model may generate:

• Inaccurate or hallucinated medical information
• Biased outputs due to limitations in training data
• Text not suitable for unsupervised clinical decision-making

Recommendations

Users should:

• Treat outputs as assistive, not authoritative
• Always cross-verify with trusted medical sources
• Avoid using the model for patient-facing diagnosis without professional review

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer, PeftModel

base_model = "openai/gpt-oss-20b"
adapter_path = "./gpt-oss-medical"

model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto", torch_dtype="auto")
model = PeftModel.from_pretrained(model, adapter_path)

tokenizer = AutoTokenizer.from_pretrained(base_model)

Training Details

Training Data

• FreedomIntelligence/medical-o1-reasoning-SFT
• Domain-specific medical text corpus
• Filtered for quality and relevance

Training Procedure

• Fine-tuned using LoRA on top of GPT-OSS-20B
• Mixed precision training (bf16)

Training Hyperparameters

• Training regime: bf16 mixed precision
• Compute: NVIDIA 2xH200 GPUs
• Training time: 12+ hours

Evaluation

Testing Data, Factors & Metrics

• Domain: medical Q&A and summarization
• Metrics: Perplexity, BLEU/ROUGE for summarization, accuracy for Q&A

Results

• [Add evaluation metrics and benchmarks]

Environmental Impact

• Hardware Type: NVIDIA 2xH200 (E2E TIR platform)
• Hours used: 12+
• Cloud Provider: E2E Networks
• Compute Region: [More Information Needed]
• Carbon Emitted: [More Information Needed]

Technical Specifications

Model Architecture and Objective

• Base: GPT-OSS-20B (20 billion parameters)
• Adapter: LoRA (low-rank fine-tuning)

Compute Infrastructure

• Hardware: NVIDIA 2xH200 (E2E TIR)
• Software: PyTorch, Transformers, PEFT

Citation

BibTeX:

@misc{gptoss20b-medical,
  title = {Medical GPT-OSS-20B LoRA Adapter},
  author = {Arunima Surendran},
  year = {2025},
  url = {https://github.com/arunimakanavu/gpt-oss-medical}
}

Model Card Authors

• Arunima Surendran

Model Card Contact

• [Your Contact Information]

Framework versions

• PEFT 0.17.0