Add IFEval reward function for instruction-following evaluation

- 112 total constraints (54 IFEval/IFTrain + 58 IFBench OOD)
- Self-contained module with no external repo dependencies
- Partial credit scoring (fraction of constraints satisfied)
- Automatic <think> tag stripping for reasoning models

Author: SandyYuan

eval_protocol/rewards/ifeval/README.md

@@ -0,0 +1,33 @@
+# IFEval Reward Function
+
+Evaluates how well model responses follow instruction constraints. Returns a partial credit score (0.0 to 1.0).
+
+## Quick Start
+
+```python
+import sys
+sys.path.insert(0, '/path/to/eval_protocol/rewards/ifeval')
+from reward import ifeval_partial_credit_reward
+
+response = "Hello world! This is my response."
+ground_truth = {
+    "instruction_id": ["keywords:existence"],
+    "kwargs": [{"keywords": ["hello", "world"]}]
+}
+
+score = ifeval_partial_credit_reward(response, ground_truth)
+# Score: 1.0 (all constraints satisfied)
+```
+
+## Dependencies
+
+```bash
+pip install spacy nltk langdetect emoji syllapy immutabledict
+python -m spacy download en_core_web_sm
+```
+
+## Notes
+
+- Automatically strips `<think>...</think>` tags before evaluation
+- Ground truth can be a dict, list, or JSON string
+- 112 total constraints (54 IFEval/IFTrain + 58 IFBench OOD)

eval_protocol/rewards/ifeval/__init__.py

@@ -0,0 +1,18 @@
+"""IFEval reward function for evaluating instruction-following capabilities.
+
+Usage:
+    # Option 1: Import spacy first to avoid cupy conflicts in some Docker environments
+    import spacy
+    from eval_protocol.rewards.ifeval import ifeval_partial_credit_reward
+
+    # Option 2: Direct import (add ifeval dir to path)
+    import sys
+    sys.path.insert(0, '/path/to/eval_protocol/rewards/ifeval')
+    from reward import ifeval_partial_credit_reward
+
+    score = ifeval_partial_credit_reward(response, ground_truth)
+"""
+
+from .reward import ifeval_partial_credit_reward
+
+__all__ = ["ifeval_partial_credit_reward"]