s-emanuilov/gemma-3-4b-rivers-abstain

Fine-Tuned Model for Epistemic Abstention

Google Gemma 3-4B-Instruct fine-tuned to abstain from answering when uncertain, implementing the R-Tuning methodology. The model includes LoRA adapter weights with rank 16 and training configuration with abstention supervision, where incorrect answers in the training data were replaced with "I don't know" responses. This approach achieved 56.7% abstention precision as reported in Section 5.2 of the paper.

Citation

@article{ackermann2025stemming,
  title={Stemming Hallucination in Language Models Using a Licensing Oracle},
  author={Ackermann, Richard and Emanuilov, Simeon},
  year={2025}
}