TrustGen/TrustEval-toolkit

TrustEval is an open-source toolkit and benchmark for evaluating the trustworthiness of generative foundation models. It covers large language models, vision-language models and text-to-image models. Many trustworthiness benchmarks use fixed test sets, which models can be tuned to and which go out of date. TrustEval instead generates new test data for each evaluation run. It was developed under the TrustGen research programme, led from the University of Notre Dame with collaborators from many universities and companies. It was presented as a demo at NAACL 2025 and at ICLR 2026.

An evaluation follows a fixed pipeline. The user downloads metadata for the chosen dimension, and the toolkit uses it to generate a test dataset. A contextual variation step then rewrites the test cases, for example by changing their wording or format, to check whether model behaviour stays consistent. The toolkit collects responses from the models under test, which can run locally or through application programming interfaces (APIs) from providers such as OpenAI, Anthropic, Google and DeepInfra.

A separate judge model, which can itself be a language, vision or image model, assesses each response. The toolkit then calculates metrics for each dimension. Results are presented in an interactive report showing each model’s scores, a written comparison with other models, examples of errors with their judgements, and a leaderboard. The toolkit is written in Python and released under the CC BY-NC 4.0 licence, which allows non-commercial use only.

Auto-discovered on 2026-10-21 by OECD Catalogue Automation