| name: IFMix | |
| description: > | |
| IFMix is an instruction-following benchmark with 400 samples across | |
| 10 categories (format lock, negative constraints, multi-turn persistent | |
| rules, system vs user conflicts, tool use, ...). | |
| evaluation_framework: "inspect-ai" | |
| tasks: | |
| - id: ifmix | |
| config: default | |
| split: train | |
| field_spec: | |
| input: messages | |
| target: constraints | |
| solvers: | |
| - name: generate | |
| scorers: | |
| - name: model_graded_qa | |
| args: | |
| model: openai/o3-mini |