This is the LACIE fine-tuned version of Llama-3 70B, finetuned according to our paper LACIE: Listener-Aware Finetuning for Confidence Calibration in Large Language Models
This model is a fine-tuned version of Llama-3 70B that has been finetuned using data from TriviaQA. LACIE is pragmatic preference-based finetuning method that optimizes models to be better calibrated w.r.t. both implicit and explicit confidence statements. The preferences in the dataset are based on correctness and whether listener accepted to rejected the answer. For more details, please see our paper.
Model Architecture
The architecture is the same as Llama-3 70B; the weights in this repo are adapter weights for Llama.