Training Language Models on the Knowledge Graph: Insights on Hallucinations and Their Detectability

TL;DR

This study investigates how the size and training duration of language models affect hallucinations, especially those verifiable from training data, revealing that larger models hallucinate less but are harder to detect.

Contribution

It introduces a knowledge graph-based dataset to control training data and systematically analyzes hallucination frequency and detectability across different model scales.

Findings

Larger models hallucinate less on fixed datasets.

Hallucinating on ≤5% of data requires significantly larger models.

Detectability of hallucinations decreases as model size increases.

Abstract

While many capabilities of language models (LMs) improve with increased training budget, the influence of scale on hallucinations is not yet fully understood. Hallucinations come in many forms, and there is no universally accepted definition. We thus focus on studying only those hallucinations where a correct answer appears verbatim in the training set. To fully control the training data content, we construct a knowledge graph (KG)-based dataset, and use it to train a set of increasingly large LMs. We find that for a fixed dataset, larger and longer-trained LMs hallucinate less. However, hallucinating on $\leq5$% of the training data requires an order of magnitude larger model, and thus an order of magnitude more compute, than Hoffmann et al. (2022) reported was optimal. Given this costliness, we study how hallucination detectors depend on scale. While we see detector size improves performance on fixed LM's outputs, we find an inverse relationship between the scale of the LM and the detectability of its hallucinations.