Tasks drawn from practice: credit analysis, compliance, claims, contract clauses.
Benchmark
Purple Bench FR: AI agents facing banking, insurance and legal work
Real tasks, written and graded by French professionals, to measure what models can really do in regulated professions.
First edition in preparation.How it is built
A reference answer and a grading rubric for each task, written by experts.
Every score is justified in writing and checked by a second reviewer.
A public sample for transparency; the rest stays private so models cannot train on it.