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