The TenkiBench dataset
The tasks behind the benchmark, served as JSON, JSONL or CSV with the methodology and licence in the same response. Ready-made citations in BibTeX, APA and MLA are on the site.
The dataset is the part of TenkiBench that is easiest to build on without running anything yourself. The endpoint delivers the tasks with an envelope stating version, methodology, licence and count, so an extract can be cited precisely instead of being pasted into a spreadsheet and losing its provenance. A smaller hold-out set is withheld to keep the tasks from leaking into training data, and the hashes for the withheld tasks are published so the extract can still be checked.
Interested in something similar?
We take on projects with concrete AI challenges where local execution or AI Act compliance are real requirements. Start with an AI assessment, or send us a line directly.