On March 19, IT Home learned from the official WeChat public account of CAICT that in order to understand the hallucination status of big models and promote the in-depth and practical application of big models, the Artificial Intelligence Institute of CAICT launched a big model hallucination test based on the previous AI Safety Benchmark evaluation work.
Big model hallucination (AI Hallucination) refers to the content that appears reasonable when the model generates content or answers questions, but is actually inconsistent with user input (fidelity hallucination) or inconsistent with facts (factual hallucination). With the widespread application of big models in key fields such as medicine and finance, the potential application risks brought by big model hallucinations are increasing, and are receiving widespread attention from the industry.
This round of hallucination testing will use large language models as the test object, covering two types of hallucinations: factual hallucinations and faithful hallucinations. The specific evaluation system is as follows:
The test data contains more than 7,000 Chinese test samples. The test format includes two types of questions: information extraction and knowledge reasoning corresponding to faithful hallucination detection, and fact discrimination questions corresponding to factual hallucination detection. In general, it involves five test dimensions: humanities, social sciences, natural sciences, applied sciences, and formal sciences.
China Academy of Information and Communications Technology invites relevant companies to participate in model evaluation and jointly promote the safe application of large models.