aydin237/medbrain-squad-auditor

MedBrainSquad — Auditor Agent Adapter

LoRA fine-tuned adapter for MedGemma 1.5 4B, trained via adversarial Supervised Fine-Tuning (SFT) to detect hallucinations, flag clinical inconsistencies, and validate SOAP JSON output from the Scribe agent.

Project

Part of the MedBrainSquad Clinical Safety Net — a three-tier adversarial multi-agent pipeline for safe, hallucination-resistant clinical documentation.

🔗 Full project: github.com/aydiny/medbrain-squad 🔗 Scribe Adapter: aydin237/medbrain-squad-scribe

Training Details

• Base model: google/medgemma-1.5-4b-it
• Method: LoRA (Rank 16, Alpha 32)
• Training approach: Adversarial SFT on synthetic hallucinated/corrupted SOAP examples
• Dataset: 500-row adversarial dataset with deliberate hallucinations and clinical errors
• Hardware: Dual NVIDIA T4 (Kaggle)
• Framework: HuggingFace PEFT + TRL

Intended Use

Research and development of adversarial safety layers in clinical NLP pipelines. Demonstrates hallucination detection in regulated healthcare environments (NHS/EU GDPR).

⚠️ Limitations & Disclaimer

• Research prototype — not validated for clinical deployment
• Base model (MedGemma 1.5 4B) subject to Google HAID Terms
• Auditor itself may miss hallucinations in novel edge cases
• Trained on synthetic data only