Reward verification and evaluation fixes (#55)

* bump up deps, fix aime24 evals, make grpo more strict

* minor fixes

* 🤨 fmt

* Update src/open_r1/grpo.py

Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.com>

---------

Co-authored-by: Hynek Kydlicek <kydlicek.hynek@huggingface.co>
Co-authored-by: Quentin Gallouédec <45557362+qgallouedec@users.noreply.github.com>

setup.py

@@ -54,7 +54,7 @@ _deps = [
     "isort>=5.12.0",
     "liger_kernel==0.5.2",
     "lighteval @ git+https://github.com/huggingface/lighteval.git@4f381b352c0e467b5870a97d41cb66b487a2c503#egg=lighteval[math]",
-    "math-verify",  # Used for math verification in grpo
+    "math-verify>=0.3.2",  # Used for math verification in grpo
     "packaging>=23.0",
     "parameterized>=0.9.0",
     "pytest",

slurm/evaluate.slurm

@@ -43,6 +43,7 @@ lighteval vllm $MODEL_ARGS "custom|$TASK|0|0" \
     --custom-tasks src/open_r1/evaluate.py \
     --use-chat-template \
     --system-prompt="Please reason step by step, and put your final answer within \boxed{}." \
+    --save-details
     --output-dir $OUTPUT_DIR 
 
 

src/open_r1/evaluate.py

@@ -24,7 +24,7 @@ from lighteval.tasks.requests import Doc
 from lighteval.utils.language import Language
 
 
-metric = multilingual_extractive_match_metric(
+latex_gold_metric = multilingual_extractive_match_metric(
     language=Language.ENGLISH,
     fallback_mode="first_match",
     precision=5,
@@ -33,6 +33,15 @@ metric = multilingual_extractive_match_metric(
     aggregation_function=max,
 )
 
+expr_gold_metric = multilingual_extractive_match_metric(
+    language=Language.ENGLISH,
+    fallback_mode="first_match",
+    precision=5,
+    gold_extraction_target=(ExprExtractionConfig(),),
+    pred_extraction_target=(ExprExtractionConfig(), LatexExtractionConfig()),
+    aggregation_function=max,
+)
+
 
 def prompt_fn(line, task_name: str = None):
     """Assumes the model is either prompted to emit \\boxed{answer} or does so automatically"""
@@ -44,11 +53,20 @@ def prompt_fn(line, task_name: str = None):
     )
 
 
+def aime_prompt_fn(line, task_name: str = None):
+    return Doc(
+        task_name=task_name,
+        query=line["problem"],
+        choices=[line["answer"]],
+        gold_index=0,
+    )
+
+
 # Define tasks
 aime24 = LightevalTaskConfig(
     name="aime24",
     suite=["custom"],
-    prompt_function=prompt_fn,
+    prompt_function=aime_prompt_fn,
     hf_repo="HuggingFaceH4/aime_2024",
     hf_subset="default",
     hf_avail_splits=["train"],
@@ -56,7 +74,7 @@ aime24 = LightevalTaskConfig(
     few_shots_split=None,
     few_shots_select=None,
     generation_size=32768,
-    metric=[metric],
+    metric=[expr_gold_metric],
     version=1,
 )
 math_500 = LightevalTaskConfig(
@@ -70,7 +88,7 @@ math_500 = LightevalTaskConfig(
     few_shots_split=None,
     few_shots_select=None,
     generation_size=32768,
-    metric=[metric],
+    metric=[latex_gold_metric],
     version=1,
 )
 

src/open_r1/grpo.py

@@ -17,7 +17,8 @@ from dataclasses import dataclass, field
 
 from datasets import load_dataset
 
-from math_verify import parse, verify
+from latex2sympy2_extended import NormalizationConfig
+from math_verify import LatexExtractionConfig, parse, verify
 from trl import GRPOConfig, GRPOTrainer, ModelConfig, ScriptArguments, TrlParser, get_peft_config
 
 
@@ -42,13 +43,36 @@ def accuracy_reward(completions, solution, **kwargs):
     contents = [completion[0]["content"] for completion in completions]
     rewards = []
     for content, sol in zip(contents, solution):
-        try:
-            answer = parse(content)
-            reward = float(verify(answer, parse(sol)))
-        except Exception:  # if it fails for any reason, return 0.0
-            reward = 0.0
+        gold_parsed = parse(sol, extraction_mode="first_match", extraction_config=[LatexExtractionConfig()])
+        if len(gold_parsed) != 0:
+            # We require the answer to be provided in correct latex (no malformed operators)
+            answer_parsed = parse(
+                content,
+                extraction_config=[
+                    LatexExtractionConfig(
+                        normalization_config=NormalizationConfig(
+                            nits=False,
+                            malformed_operators=False,
+                            basic_latex=True,
+                            equations=True,
+                            boxed=True,
+                            units=True,
+                        ),
+                        # Ensures that boxed is tried first
+                        boxed_match_priority=0,
+                        try_extract_without_anchor=False,
+                    )
+                ],
+                extraction_mode="first_match",
+            )
+            # Reward 1 if the content is the same as the ground truth, 0 otherwise
+            reward = float(verify(answer_parsed, gold_parsed))
+        else:
+            # If the gold solution is not parseable, we reward 1 to skip this example
+            reward = 1.0
+            print("Failed to parse gold solution: ", sol)
         rewards.append(reward)
-    # Reward 1 if the content is the same as the ground truth, 0 otherwise
+
     return rewards