Good benchmarking practices

Informative and powerful benchmarks

Developers of GPAI systems can select benchmarks that are difficult enough to be informative about the capabilities of their AI systems, and cover a large spectrum of domains in order to signal areas where the GPAI system is performing poorly [120]. Suitable benchmarks contain no label errors, are not vulnerable to being benchmark contaminants, and are often audited by independent domain experts if they contain domain-specific questions. For multimodal GPAI systems, good benchmarks cover every modality, especially the interaction of different modalities.

Benchmark dataset auditing

Auditing benchmark datasets allows for verification of the utility and limitations of the datasets [158]. This allows the provider to more accurately measure AI model capabilities and safety. Auditing includes the evaluation of such datasets by independent third-party organizations and the release of benchmark dataset metadata to the auditors.

Dynamic benchmarking

Dynamic benchmarks are benchmarks that can be continuously updated with new human-generated data. By having one or more target models “in the loop,” examples for benchmarking can be generated with the intent of fooling these target models, or to assess if these models express an appropriate level of uncertainty [103]. As the dataset in the benchmark grows, previously benchmarked models can also be reassessed against the updated dataset to reflect its performance in a more representative manner.