Math extraction - allow only trying the first match, more customizable latex extraction + bump deps (#522)

* extract matching

* better docstring

* lazy imports

* bump up math

* Update src/lighteval/metrics/dynamic_metrics.py

Co-authored-by: Clémentine Fourrier <[email protected]>

* fix pr commnets

* Apply suggestions from code review

Co-authored-by: Clémentine Fourrier <[email protected]>

* rename comparisson -> comparison

* fix expr numbers extraction with currency or units

* add test for correct extraction of failed answer

* bump of latex2sympy2 version, add new tests for extract metric

* bump up latex2sympy + adjust latex target

* revert gold target timoeut

* remove dead comment 💀

* add doc

---------

Co-authored-by: Clémentine Fourrier <[email protected]>

Author: hynky1999

pyproject.toml

@@ -109,7 +109,7 @@ multilingual = [
     "jieba", # for chinese tokenizer
     "pyvi", # for vietnamese tokenizer
 ]
-math = ["latex2sympy2_extended>=0.9.1"]
+math = ["latex2sympy2_extended>=0.9.3"]
 
 [project.urls]
 Homepage = "https://github.com/huggingface/lighteval"

src/lighteval/metrics/dynamic_metrics.py

@@ -191,6 +191,7 @@ def multilingual_extractive_match_metric(
     pred_extraction_target: Sequence[ExtractionTarget] = (ExprExtractionConfig(),),
     aggregation_function: Callable[[list[float]], float] = max,
     fallback_mode: Literal["no_fallback", "first_match"] = "first_match",
+    extraction_mode: Literal["first_match", "any_match"] = "any_match",
     precision: int = 6,
 ) -> SampleLevelMetric:
     """Creates a language-aware extractive match metric that extracts answers from the model's output.
@@ -215,6 +216,10 @@ def multilingual_extractive_match_metric(
             How to perform extraction. Defaults to "first_match".
             - "no_fallback": Only use first successfully parsed matches
             - "first_match": Use the first successfully parsed match + first match irregardless the parsing success
+        extraction_mode: Literal["first_match", "any_match"]
+            - "first_match": Only tries to extract the first regex match if it fails no other matches are tried
+            - "any_match": Tries to extract any regex match
+
         precision: int
             Number of decimal places to use when comparing numerical values. Defaults to 6.
 
@@ -240,9 +245,12 @@ def sample_level_fn(golds: list[str], predictions: list[str], formatted_doc: Doc
         pred_extraction_regexes = get_extraction_regexes(formatted_doc, pred_extraction_target, language)
 
         extracted_predictions = [
-            extract_target_from_pred(pred, pred_extraction_regexes, fallback_mode) for pred in predictions
+            extract_target_from_pred(pred, pred_extraction_regexes, fallback_mode, extraction_mode)
+            for pred in predictions
+        ]
+        extracted_golds = [
+            extract_target_from_pred(gold, gold_extraction_regexes, fallback_mode, extraction_mode) for gold in golds
         ]
-        extracted_golds = [extract_target_from_pred(gold, gold_extraction_regexes, fallback_mode) for gold in golds]
 
         # Assert on empty gold and warn on empty pred
         if any(len(g) == 0 for g in extracted_golds):

src/lighteval/metrics/utils/extractive_match_utils.py

@@ -21,10 +21,10 @@
 # SOFTWARE.
 
 import re
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from functools import lru_cache
 from itertools import groupby
-from typing import Literal, Sequence
+from typing import Any, Literal, Sequence
 
 import sympy
 from sympy import Basic, MatrixBase, Number
@@ -39,17 +39,33 @@
 from lighteval.utils.timeout import timeout
 
 
+@requires_latex2sympy2_extended
+def latex_normalization_config_default_factory():
+    from latex2sympy2_extended.latex2sympy2 import NormalizationConfig
+
+    return NormalizationConfig(
+        basic_latex=True,
+        units=True,
+        malformed_operators=True,
+        nits=True,
+        boxed=True,
+        equations=True,
+    )
+
+
 @dataclass(frozen=True)
 class LatexExtractionConfig:
     """Config for extracting latex from the prediction.
 
     Attributes:
         try_extract_without_anchor (bool): Whether to try extracting latex without requiring specific anchors like "answer:" or "final answer is"
-        enforce_boxed_match (bool): Whether to also consider extracting from plain \boxed{...} expressions
+        boxed_match_priority (int): Priority of the boxed match regex (-1 never, 0 first, 55 after final answer: anchor, etc...)
+        normalization_config (latex2sympy2_extended.latex2sympy2.NormalizationConfig): Normalization config to use for latex extraction
     """
 
     try_extract_without_anchor: bool = True
-    enforce_boxed_match: bool = True
+    boxed_match_priority: int = 55
+    normalization_config: Any = field(default_factory=latex_normalization_config_default_factory)
 
 
 @dataclass(frozen=True)
@@ -187,9 +203,8 @@ def lazy_latex_regex(latex_config: LatexExtractionConfig, language: Language) ->
         if latex_config.try_extract_without_anchor:
             regexes.append((latex_re, 300))
 
-    # This ensures that boxed is matched right after the final answer xxxx
-    if latex_config.enforce_boxed_match:
-        regexes.append((latex_boxed, 55))
+    if latex_config.boxed_match_priority >= 0:
+        regexes.append((latex_boxed, latex_config.boxed_match_priority))
 
     return [(re.compile(pattern, re.DOTALL), priority) for pattern, priority in regexes]
 
@@ -387,6 +402,7 @@ def extract_target_from_pred(
     pred: str,
     target_res: list[tuple[list[tuple[re.Pattern[str], int]], ExtractionTarget]],
     fallback_mode: Literal["no_fallback", "first_match"] = "no_fallback",
+    extraction_mode: Literal["first_match", "any_match"] = "any_match",
 ):
     """Extracts targets from a prediction string using regex patterns.
     Returns first sucesffuly extracted match.
@@ -397,6 +413,9 @@ def extract_target_from_pred(
         fallback_mode (Literal["no_fallback", "first_match"], optional): How to handle extraction failures. Defaults to "no_fallback".
             - "no_fallback": Return only successfully parsed match
             - "first_match": Additionaly Include the first string match no matter how parsing finished
+        extraction_mode (Literal["first_match", "any_match"], optional): How to handle extraction failures. Defaults to "any_match".
+            - "first_match": Only tries to extract the first match
+            - "any_match": Tries to extract any match
 
     Returns:
         list: List of extracted predictions, with first fallbac string appended if fallback_mode is "first_match"
@@ -410,6 +429,7 @@ def extract_target_from_pred(
         for target_patterns, target_type in target_res
         for pattern, priority in target_patterns
     ]
+    match_found = False
 
     # Group patterns by priority using itertools.groupby
     for _, patterns_group in groupby(sorted(all_patterns, key=lambda x: x[2]), key=lambda x: x[2]):
@@ -426,6 +446,7 @@ def extract_target_from_pred(
         # Try to extract from each match, starting from rightmost
         for match, _, _, target_type in matches_with_pos:
             extracted_match, str_fallback = extract_match(match, target_type)
+            match_found = True
 
             if str_fallback:
                 fallbacks.append(str_fallback)
@@ -434,8 +455,11 @@ def extract_target_from_pred(
                 extracted_predictions.append(extracted_match)
                 break
 
+            if extraction_mode == "first_match":
+                break
+
         # If we found something and we're in first_match mode, stop processing other priorities
-        if extracted_predictions:
+        if extracted_predictions or (match_found and extraction_mode == "first_match"):
             break
 
     if fallback_mode == "first_match" and fallbacks: