Base model: google/medgemma-4b-it
Method: RLFR with 6 reward features (factual accuracy, completeness, format compliance, etc.)
Task: Generate MedWatch 3500A pharmacovigilance narratives from CRF data

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("AlphaRaven/medgemma-4b-antihallu")
tokenizer = AutoTokenizer.from_pretrained("AlphaRaven/medgemma-4b-antihallu")

Part of the Clinical Trial Simulation Engine pipeline.