Overview

This project studies whether a small LLM can be trained to express uncertainty instead of confidently hallucinating. The model is fine-tuned to prefer safe abstention when it lacks reliable knowledge.

Key Highlights

  • Fine-tunes Llama 3.2 1B Instruct on an uncertainty-focused instruction dataset.
  • Demonstrates a qualitative reduction in hallucinated answers on unknown questions.
  • Shows an important trade-off: over-abstention, where the model may say “I don’t know” even for answerable prompts.
  • Includes links to both the model and dataset on Hugging Face.

Technologies

  • Core: Python, PyTorch, Hugging Face ecosystem
  • Training Setup: Google Colab (T4 GPU)
  • Approach: Instruction tuning focused on uncertainty-aware responses

View Code on GitHub