RKB109/clinical-rag-safety-gateway-20260815-model

Clinical RAG Safety Gateway Baseline Model

Model Description

This repository contains a small, transparent prototype model for Clinical assistants need retrieval, source attribution, and explicit abstention before answers reach care teams.

The model combines per-label token weights with IDF-weighted evidence retrieval. It was generated for reproducible architecture demonstrations and does not call a hosted LLM.

Evaluation

• Held-out synthetic examples: 4
• Accuracy: 1
• Intended metrics: retrieval_accuracy, abstention_coverage, citation_coverage

Intended Use

• Architecture prototyping
• CI and evaluation examples
• Local baseline comparisons
• Educational experimentation

Hugging Face Task Coverage

• question-answering
• sentence-similarity
• text-classification
• summarization

Limitations and Risks

Synthetic educational data only. The baseline must not provide diagnosis, treatment, or emergency medical advice.

The dataset is synthetic and small. Do not use this model for consequential decisions without representative data, expert review, and production-grade evaluation.

Reproducibility

The linked GitHub repository includes train.py, the exact dataset split, evaluation code, and the model JSON format.