| --- |
| license: mit |
| task_categories: |
| - text-generation |
| language: |
| - en |
| pretty_name: IF Mix |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train.jsonl |
| tags: |
| - benchmark |
| - instruction-following |
| --- |
| |
| # IFMix |
|
|
| IFMix is an instruction-following benchmark with 400 samples across 10 categories: |
| format lock, lexical constraints, negative constraints, minimal output, |
| multi-step procedure, transformation tasks, persistent rules (multi-turn), |
| system vs user, conflict resolution, and tool use. |
|
|
| ## Fields |
|
|
| - `id`: unique sample id |
| - `category`: one of the 10 categories |
| - `messages`: conversation (list of `{role, content}`) |
| - `constraints`: natural-language criteria a correct reply must satisfy |
|
|
| ## Evaluation |
|
|
| Each constraint is checked with an LLM judge (see `eval.yaml`). |
