Benchmark
FlavourBench
Evaluate ingredient substitution, pairing, and constrained culinary choices with executable scoring.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
The revised paper evaluates 27 model endpoints on 534 ingredient-selection tasks.
Methodology
Compile deterministic answer maps from a versioned culinary embedding model. Score ingredient portfolios while preserving reward-model and task-panel versions.