Update eval.yaml

eval.yaml

@@ -0,0 +1,22 @@
+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, ...).
+
+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