Bio-GPT-OSS-20B

A specialized biomedical question-answering model fine-tuned from gpt-oss-20B on scholarly Q&A extraction tasks from biomedical research papers.

Bio-GPT-OSS-20B/ Bio-GPT-OSS-20B pubmed

Model Details

• Base Model: gpt-oss-20B
• Fine-tuned on: Biomedical scholarly Q&A dataset
• Task: Question answering from scientific papers with step-by-step reasoning
• Parameters: ~20B
• Language: English

Training Data

The model was trained on ~ 1000 conversation pairs extracted from biomedical research papers, featuring:

• Dataset: Pubmed-Bio-Thinking-1000
• Data: 100000 & 1000000 version of the Dataset is avalible, contact for access
• System: Scholarly Q&A extraction agent instructions
• User: Complex biomedical questions about research findings
• Analysis: Detailed step-by-step reasoning and evidence extraction
• Final: Concise, precise answers with citations

Data Structure

{
  "paper_id": "PMC8026465",
  "reasoning_language": "English",
  "developer": "You are a scholarly Q & A extraction agent. Only use the provided paper. Be concise and precise.",
  "user": "According to the comparative analysis by Wu et al., how do the kidney organoid differentiation protocols compare?",
  "analysis": "Steps:\n1) Locate the section \"3D kidney organoids\"...\n2) Find Wu et al. comparative study...",
  "final": "According to Wu et al., both Morizane and Takasato protocols generated immature tissue, expressing ~20% of adult transcription factors.",
  "messages": [
    {"role": "system", "content": "...", "thinking": null},
    {"role": "user", "content": "...", "thinking": null},
    {"role": "assistant", "content": "...", "thinking": "Steps:\n1) Locate..."}
  ]
}

Sample Conversations

Question: Kidney organoid protocol comparison Analysis: Multi-step evidence extraction from specific paper sections Answer: Quantitative comparison with statistical details (both protocols ~20% maturity)

Question: Cognitive test predictive ability differences
Analysis: Statistical result extraction and comparison across test conditions Answer: 1-week test superior predictive power (p=.003 vs p=.11)

Intended Use

• Biomedical research question answering
• Scientific paper analysis and extraction
• Evidence-based reasoning in life sciences
• Academic research assistance

Limitations

• Domain-specific to biomedical/life sciences literature
• Requires access to source papers for optimal performance
• May not generalize well outside scholarly contexts

Citation

@model{bio-gpt-oss-20b,
  title={Bio-GPT-OSS-20B: Biomedical Question Answering with Reasoning},
  author={Your Name},
  year={2024},
  base_model={gpt-oss-20B}
}

This gpt_oss model was trained 2x faster with Unsloth and Huggingface's TRL library.