Add git diff reward

pyproject.toml

@@ -36,6 +36,7 @@ dependencies = [
     "reasoning-gym @ git+https://github.com/open-thought/reasoning-gym.git",
     "tomli>=2.2.1",
     "pydantic[email]>=2.11.5",
+    "cydifflib>=1.2.0",
 ] 
 
 [project.optional-dependencies]

src/zeroband/inference/genesys/__init__.py

@@ -4,6 +4,7 @@
 from zeroband.inference.genesys.code import evaluate_code
 from zeroband.inference.genesys.code_output_prediction import verify_code_output_prediction
 from zeroband.inference.genesys.complex_json_output import verify_complex_json_formatting
+from zeroband.inference.genesys.git_diff import compute_git_diff_reward
 from zeroband.inference.genesys.ifeval import verify_ifeval
 from zeroband.inference.genesys.math import compute_math_reward
 from zeroband.inference.genesys.pydantic_json_adherance import validate_pydantic_json
@@ -22,6 +23,7 @@
     "pydantic_adherance",
     "ifeval",
     "complex_json_output",
+    "git_diff",
 ]
 
 
@@ -43,4 +45,5 @@ def get_reward_function(task_type: TaskType) -> Callable[[str, dict], float]:
     "pydantic_adherance": validate_pydantic_json,
     "ifeval": verify_ifeval,
     "complex_json_output": verify_complex_json_formatting,
+    "git_diff": compute_git_diff_reward,
 }

src/zeroband/inference/genesys/git_diff.py

@@ -0,0 +1,101 @@
+# Acknowledgements:
+#   SWE-Fixer: Training Open-Source LLMs for Effective and Efficient GitHub Issue Resolution
+#   Xie, Chengxing et al., 2025
+#
+#   Agentless: Demystifying LLM-based Software Engineering Agents
+#   Xia, Chunqiu Steven et al., 2024
+#
+#   SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution
+#   Yuxiang Wei et al., 2025
+
+import re
+from typing import Dict
+
+import cydifflib
+
+DIFF_BLOCK_REGEX = re.compile(r"```diff\s*(.*?)\s*```", re.DOTALL)
+INDEX_LINE_REGEX = re.compile(r"^index [^\n]*\n")
+FUNC_CONTEXT_REGEX = re.compile(r"(?m)^(@@[^@]*@@).*")
+
+
+def parse_last_diff_codeblock(markdown_str: str) -> str:
+    """Extract the last ```diff``` code block from markdown text."""
+    matches = DIFF_BLOCK_REGEX.findall(markdown_str)
+    if matches:
+        return matches[-1].strip()
+    else:
+        return ""
+
+
+def normalize_diff(diff_text: str) -> str:
+    """
+    Normalize diff text by removing lines starting with 'index ...' and stripping function context after @@.
+    The function context/section header can differ between diffs, because language specific parsing might not be enabled.
+
+    Example:
+    ```diff
+    diff --git a/file.py b/file.py
+    index 1234567890..1234567890
+    --- a/file.py
+    +++ b/file.py
+    @@ -15,1 +15,1 @@ def some_func():
+    -    pass
+    +    return
+    ```
+
+    becomes:
+    ```diff
+    diff --git a/file.py b/file.py
+    --- a/file.py
+    +++ b/file.py
+    @@ -15,1 +15,1 @@
+    -    pass
+    +    return
+    ```
+    """
+    diff_text = INDEX_LINE_REGEX.sub("", diff_text)
+    diff_text = FUNC_CONTEXT_REGEX.sub(r"\1", diff_text)
+    diff_text = diff_text.strip() + "\n"
+    return diff_text
+
+
+def compute_git_diff_reward(completion: str, verification_info: Dict) -> float:
+    """
+    Compute reward for git diff generation tasks using LCS (Longest Common Subsequence) ratio.
+
+    Args:
+        completion: Model's response string
+        verification_info: Dict containing golden_diff
+
+    Returns:
+        Float score (0.0 to 1.0) representing diff similarity
+    """
+    # Extract the response after thinking (if present)
+    if "</think>" in completion:
+        response = completion.split("</think>")[1].strip()
+    else:
+        response = completion.strip()
+
+    if not response:
+        return 0.0
+
+    # Get expected answer from verification_info
+    golden_diff = verification_info.get("golden_diff", "")
+    if not golden_diff:
+        return 0.0
+
+    try:
+        # Extract diff from response
+        response_diff = parse_last_diff_codeblock(response)
+        response_diff = normalize_diff(response_diff)
+
+        if not response_diff.strip():
+            return 0.0
+
+        # Calculate LCS ratio
+        similarity = cydifflib.SequenceMatcher(None, response_diff, golden_diff, autojunk=False).ratio()
+
+        return similarity
+
+    except Exception:
+        return 0.0