format-bench
Data-format conversion benchmark

Can a model rewrite your data without breaking it?

Converting between text, JSON, YAML and CSV looks trivial and isn't. A model has to keep every field, infer the types CSV threw away, strip units, clamp ranges, escape correctly — and emit the target format and nothing else. This benchmark measures exactly that, on 2,025 frozen cases across nine directions, with every model run twice: once with plain instructions, once with a worked example.

models
2,025test cases
9directions
7constraints
best task score
loading results…