Merge branch 'huggingface:main' into add_swiss_legal_evals

Author: rolshoven

README.md

@@ -21,7 +21,7 @@
 
 ---
 
-**Documentation**: <a href="https://github.com/huggingface/lighteval/wiki" target="_blank">Lighteval's Wiki</a>
+**Documentation**: <a href="https://huggingface.co/docs/lighteval/index" target="_blank">Lighteval's Wiki</a>
 
 ---
 

pyproject.toml

@@ -95,7 +95,7 @@ tensorboardX = ["tensorboardX"]
 vllm = ["vllm", "ray", "more_itertools"]
 quality = ["ruff==v0.2.2","pre-commit"]
 tests = ["pytest==7.4.0"]
-dev = ["lighteval[accelerate,quality,tests,multilingual]"]
+dev = ["lighteval[accelerate,quality,tests,multilingual,math]"]
 docs = ["hf-doc-builder", "watchdog"]
 extended_tasks = [
   "langdetect", # ifeval
@@ -111,6 +111,7 @@ multilingual = [
     "jieba", # for chinese tokenizer
     "pyvi", # for vietnamese tokenizer
 ]
+math = ["latex2sympy2_extended>=0.9.0"]
 
 [project.urls]
 Homepage = "https://github.com/huggingface/lighteval"

src/lighteval/data.py

@@ -25,8 +25,15 @@
 from typing import Iterator, Tuple
 
 import torch
+from packaging import version
 from torch.utils.data import Dataset
-from torch.utils.data.distributed import DistributedSampler, T_co
+
+
+if version.parse(torch.__version__) >= version.parse("2.5.0"):
+    from torch.utils.data.distributed import DistributedSampler, _T_co
+else:
+    from torch.utils.data.distributed import DistributedSampler
+    from torch.utils.data.distributed import T_co as _T_co
 
 from lighteval.tasks.requests import (
     GreedyUntilRequest,
@@ -318,7 +325,7 @@ class GenDistributedSampler(DistributedSampler):
     as our samples are sorted by length.
     """
 
-    def __iter__(self) -> Iterator[T_co]:
+    def __iter__(self) -> Iterator[_T_co]:
         if self.shuffle:
             # deterministically shuffle based on epoch and seed
             g = torch.Generator()

src/lighteval/metrics/dynamic_metrics.py

@@ -20,7 +20,8 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 # SOFTWARE.
 
-from typing import Callable, Literal
+import logging
+from typing import Callable, Literal, Sequence
 
 import numpy as np
 
@@ -37,8 +38,22 @@
     LogProbTokenNorm,
     get_multilingual_normalizer,
 )
+from lighteval.metrics.utils.extractive_match_utils import (  # noqa: F401
+    ExprExtractionConfig,
+    ExtractionTarget,
+    IndicesExtractionConfig,
+    LatexExtractionConfig,
+    extract_target_from_pred,
+    get_extraction_regexes,
+)
+from lighteval.metrics.utils.math_comparison import compare_gold_target
 from lighteval.metrics.utils.metric_utils import MetricCategory, MetricUseCase, SampleLevelMetric
+from lighteval.tasks.requests import Doc
 from lighteval.utils.language import Language
+from lighteval.utils.timeout import timeout
+
+
+logger = logging.getLogger(__name__)
 
 
 def loglikelihood_acc_metric(normalization: LogProbNormalization | None = None) -> SampleLevelMetric:
@@ -168,3 +183,94 @@ def multilingual_quasi_exact_match_metric(
         corpus_level_fn=np.mean,
         higher_is_better=True,
     )
+
+
+def multilingual_extractive_match_metric(
+    language: Language = Language.ENGLISH,
+    gold_extraction_target: Sequence[ExtractionTarget] = (ExprExtractionConfig(),),
+    pred_extraction_target: Sequence[ExtractionTarget] = (ExprExtractionConfig(),),
+    aggregation_function: Callable[[list[float]], float] = max,
+    fallback_mode: Literal["no_fallback", "first_match"] = "first_match",
+    precision: int = 6,
+) -> SampleLevelMetric:
+    """Creates a language-aware extractive match metric that extracts answers from the model's output.
+
+    Known issues:
+    - If the task is to simplify an expression, the metric might overestimate the accuracy. This is because if the model doesn't output any anchor for the extraction (e.g final answer is..),
+        it's possible that the the extracted prediction will be the expression to simplify. Because we do simplifications ourselves, it can thus happen that sympy will correctly simplify the expression,
+        thus it will match gold, despite model not doing anything. PRs to fix this are welcome.
+
+    - There is currently no StringExtractionConfig, so if the gold is \boxed{\text{Friday}} and model outputs Friday it will not match, because nothing will be extracted.
+
+    Args:
+        language: Language
+            The language of the samples.
+        gold_extraction_target: Sequence[ExtractionTarget]
+            Extraction targets to use for gold answers. Defaults to extracting simple math expressions.
+        pred_extraction_target: Sequence[ExtractionTarget]
+            Extraction targets to use for predictions. Defaults to extracting simple math expressions.
+        aggregation_function: Callable[[list[float]], float]
+            Function to aggregate scores when multiple golds/predictions are present. Defaults to max.
+        fallback_mode: Literal["no_fallback", "first_match"]
+            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
+        precision: int
+            Number of decimal places to use when comparing numerical values. Defaults to 6.
+
+    Returns:
+        A sample level metric that extracts and compares mathematical expressions.
+
+    """
+
+    @timeout(2)
+    def add_to_specifics_with_timeout(
+        formatted_doc: Doc, extracted_predictions: list[list[str]], extracted_golds: list[list[str]]
+    ) -> None:
+        if formatted_doc.specific is None:
+            formatted_doc.specific = {}
+
+        formatted_doc.specific["extracted_predictions"] = [
+            str(pred) for preds in extracted_predictions for pred in preds
+        ]
+        formatted_doc.specific["extracted_golds"] = [str(gold) for golds in extracted_golds for gold in golds]
+
+    def sample_level_fn(golds: list[str], predictions: list[str], formatted_doc: Doc) -> float:
+        gold_extraction_regexes = get_extraction_regexes(formatted_doc, gold_extraction_target, language)
+        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
+        ]
+        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):
+            raise ValueError(f"No gold targets found for at least one gold. Gold: {golds}, Pred: {predictions}")
+
+        if all(len(p) == 0 for p in extracted_predictions):
+            logger.warning(
+                f"We did not manage to extract a prediction in the correct format. Gold: {golds}, Pred: {predictions}"
+            )
+
+        # We have to use timeout because the sypmy to str conversion can be very slow
+        try:
+            add_to_specifics_with_timeout(formatted_doc, extracted_predictions, extracted_golds)
+        except:  # noqa: E722
+            logger.warning("Timeout when adding extracted predictions and golds to specific")
+
+        return aggregation_function(
+            [
+                (1.0 if any(compare_gold_target(gold, pred, precision) for gold in extracted_golds) else 0.0)
+                for pred in extracted_predictions
+            ]
+        )
+
+    return SampleLevelMetric(
+        metric_name="extractive_match",
+        sample_level_fn=sample_level_fn,
+        category=MetricCategory.GENERATIVE,
+        use_case=MetricUseCase.ACCURACY,
+        corpus_level_fn=np.mean,
+        higher_is_better=True,
+    )

src/lighteval/metrics/llm_as_judge.py

@@ -193,22 +193,32 @@ def __call_litellm(self, prompts):
         import litellm
 
         def __call_api(prompt):
+            error_message = "ERROR: Failed to get response from the API."
             for _ in range(self.API_MAX_RETRY):
                 try:
-                    response = litellm.completion(
-                        model=self.model,
-                        messages=prompt,
-                        response_format={"type": "text"},
-                        max_tokens=512,
-                        n=1,
-                        caching=True,
-                    )
+                    kwargs = {
+                        "model": self.model,
+                        "messages": prompt,
+                        "response_format": {"type": "text"},
+                        "max_tokens": 512,
+                        "n": 1,
+                        "caching": True,
+                    }
+                    response = litellm.completion(**kwargs)
                     text = response.choices[0].message.content
+                    if not text or text == error_message:
+                        kwargs["caching"] = False
+                        response = litellm.completion(**kwargs)
+                        text = response.choices[0].message.content
+                        if not text or text == error_message:
+                            # Just return an error response if the second attempt fails too
+                            logger.error(f"Failed to get response from the API for prompt: {prompt}")
+                            return error_message
                     return text
                 except Exception as e:
                     logger.warning(f"{type(e), e}")
                     time.sleep(self.API_RETRY_SLEEP)
-            raise Exception("Failed to get response from the API")
+            return error_message
 
         results = []
         with ThreadPoolExecutor(100) as executor:

src/lighteval/metrics/metrics_corpus.py

@@ -26,6 +26,7 @@
 """
 import logging
 import math
+from typing import Literal
 
 import numpy as np
 import sacrebleu
@@ -89,33 +90,38 @@ def compute(self, items: list[LogprobCorpusMetricInput]):
 
 
 class CorpusLevelTranslationMetric:
-    def __init__(self, metric_type: str):
+    def __init__(self, metric_type: str, lang: Literal["zh", "ja", "ko", ""] = ""):
         """Stores the relevant parameters for a corpus level translation metric.
 
         Args:
             metric_type (str): Can be any of bleu, chrf, or ter depending on the metric to use.
         """
-        if metric_type == "bleu":
-            self.metric = sacrebleu.corpus_bleu
-        elif metric_type == "chrf":
-            self.metric = sacrebleu.corpus_chrf
-        elif metric_type == "ter":
-            self.metric = sacrebleu.corpus_ter
+        self.metric_type = metric_type
+        self.lang = lang
+
+    def get_metric(self):
+        if self.metric_type == "bleu":
+            return sacrebleu.BLEU(trg_lang=self.lang)
+        elif self.metric_type == "chrf":
+            return sacrebleu.CHRF()
+        elif self.metric_type == "ter":
+            return sacrebleu.TER(asian_support=True if self.lang != "" else False)
         else:
-            raise ValueError(f"Unknown corpus level translation metric type : {metric_type}")
+            raise ValueError(f"Unknown corpus level translation metric type : {self.metric_type}")
 
     def compute(self, items: list[GenerativeCorpusMetricInput]) -> float:
         """Computes the metric score over all the corpus generated items, by using the sacrebleu implementation."""
+        metric = self.get_metric()
         golds = [i.golds for i in items]
         preds = []
         for i in items:
             pred = as_list(i.preds)
             if len(pred) > 1:
                 logger.info(
-                    f"Multiple predictions present, keeping only the first prediction (when computing sacrebleu.{self.metric.__name__})."
+                    f"Multiple predictions present, keeping only the first prediction (when computing sacrebleu.{metric.__name__})."
                 )
             preds.append(pred[0])
-        return float(self.metric(hypotheses=preds, references=golds).score)
+        return float(metric.corpus_score(hypotheses=preds, references=golds).score)
 
 
 class CorpusLevelPerplexityMetric: