Characterizing Truthfulness in Large Language Model Generations with Local Intrinsic Dimension

Authors: Fan Yin, Jayanth Srinivasa, Kai-Wei Chang

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through experiments on four question answering (QA) datasets, we demonstrate the effectiveness of our proposed method.
Researcher Affiliation Collaboration 1Department of Computer Science, University of California, Los Angeles, LA, U.S.A. 2Cisco Research, U.S.A.
Pseudocode No The paper describes the method verbally and with equations but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Code is available at: https://github.com/fanyin3639/ LID-Hallucination Detection.
Open Datasets Yes We consider four generative QA tasks: Trivia QA (Joshi et al., 2017), Co QA (Reddy et al., 2019), Hotpot QA (Yang et al., 2018), and Tydi QA-GP (English) (Clark et al., 2020).
Dataset Splits Yes For each of the datasets, we generate outputs for 2,000 samples from the validation sets and test the methods with those samples.
Hardware Specification No The paper does not provide specific details on the hardware used for experiments, such as GPU or CPU models.
Software Dependencies No The paper does not provide specific software dependency versions (e.g., Python 3.x, PyTorch 1.x) for reproducibility.
Experiment Setup Yes For LID-MLE and LID-Geo MLE, we use 500 nearest neighbors when estimating LIDs for all datasets. We follow Kuhn et al. (2022), and set the temperature to be 0.5 and the number of generated samples to be 10. We fine-tune a Llama-2-7B model for 3,000 steps, roughly 3 epochs, on SUPER-NI s training set.