Out-of-Domain Finetuning to Bootstrap Hallucination Detection
Eugene Yan
Researchers demonstrated that pre-finetuning a model on out-of-domain Wikipedia summaries before finetuning on the Factual Inconsistency Benchmark (FIB) dataset significantly improves hallucination detection in news summaries. The approach achieved a PR AUC of 0.85, representing a 23% improvement over finetuning on FIB data alone. This finding suggests that transfer learning through multi-stage finetuning can reduce the need for large amounts of task-specific labeled data by leveraging related datasets from different domains.
Why it matters
How to use open-source, permissive-use data and collect less labeled samples for our tasks.
