Merge branch 'main' into Document-Custom-Model-Files

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/logging/evaluation_tracker.py

@@ -34,7 +34,6 @@
 import torch
 from datasets import Dataset, load_dataset
 from datasets.utils.metadata import MetadataConfigs
-from fsspec import url_to_fs
 from huggingface_hub import DatasetCard, DatasetCardData, HfApi, HFSummaryWriter, hf_hub_url
 
 from lighteval.logging.info_loggers import (
@@ -53,6 +52,11 @@
 if is_nanotron_available():
     from nanotron.config import GeneralArgs  # type: ignore
 
+try:
+    from fsspec import url_to_fs
+except ImportError:
+    from fsspec.core import url_to_fs
+
 
 class EnhancedJSONEncoder(json.JSONEncoder):
     """
@@ -231,9 +235,45 @@ def save_results(self, date_id: str, results_dict: dict):
         with self.fs.open(output_results_file, "w") as f:
             f.write(json.dumps(results_dict, cls=EnhancedJSONEncoder, indent=2, ensure_ascii=False))
 
-    def save_details(self, date_id: str, details_datasets: dict[str, Dataset]):
+    def _get_details_sub_folder(self, date_id: str):
         output_dir_details = Path(self.output_dir) / "details" / self.general_config_logger.model_name
-        output_dir_details_sub_folder = output_dir_details / date_id
+        if date_id in ["first", "last"]:
+            # Get all folders in output_dir_details
+            if not self.fs.exists(output_dir_details):
+                raise FileNotFoundError(f"Details directory {output_dir_details} does not exist")
+
+            # List all folders and filter out files
+            folders = [f["name"] for f in self.fs.listdir(output_dir_details) if f["type"] == "directory"]
+
+            if not folders:
+                raise FileNotFoundError(f"No timestamp folders found in {output_dir_details}")
+
+            # Parse timestamps and get first or last
+            date_id = max(folders) if date_id == "last" else min(folders)
+        return output_dir_details / date_id
+
+    def load_details_datasets(self, date_id: str, task_names: list[str]) -> dict[str, Dataset]:
+        output_dir_details_sub_folder = self._get_details_sub_folder(date_id)
+        logger.info(f"Loading details from {output_dir_details_sub_folder}")
+        date_id = output_dir_details_sub_folder.name  # Overwrite date_id in case of latest
+        details_datasets = {}
+        for file in self.fs.glob(str(output_dir_details_sub_folder / f"details_*_{date_id}.parquet")):
+            task_name = Path(file).stem.replace("details_", "").replace(f"_{date_id}", "")
+            if "|".join(task_name.split("|")[:-1]) not in task_names:
+                logger.info(f"Skipping {task_name} because it is not in the task_names list")
+                continue
+            dataset = load_dataset("parquet", data_files=file, split="train")
+            details_datasets[task_name] = dataset
+
+        for task_name in task_names:
+            if not any(task_name.startswith(task_name) for task_name in details_datasets.keys()):
+                raise ValueError(
+                    f"Task {task_name} not found in details datasets. Check the tasks to be evaluated or the date_id used to load the details ({date_id})."
+                )
+        return details_datasets
+
+    def save_details(self, date_id: str, details_datasets: dict[str, Dataset]):
+        output_dir_details_sub_folder = self._get_details_sub_folder(date_id)
         self.fs.mkdirs(output_dir_details_sub_folder, exist_ok=True)
         logger.info(f"Saving details to {output_dir_details_sub_folder}")
         for task_name, dataset in details_datasets.items():

src/lighteval/main_accelerate.py

@@ -67,6 +67,9 @@ def accelerate(  # noqa C901
     num_fewshot_seeds: Annotated[
         int, Option(help="Number of seeds to use for few-shot evaluation.", rich_help_panel=HELP_PANEL_NAME_1)
     ] = 1,
+    load_responses_from_details_date_id: Annotated[
+        Optional[str], Option(help="Load responses from details directory.", rich_help_panel=HELP_PANEL_NAME_1)
+    ] = None,
     # === saving ===
     output_dir: Annotated[
         str, Option(help="Output directory for evaluation results.", rich_help_panel=HELP_PANEL_NAME_2)
@@ -137,6 +140,7 @@ def accelerate(  # noqa C901
         max_samples=max_samples,
         use_chat_template=use_chat_template,
         system_prompt=system_prompt,
+        load_responses_from_details_date_id=load_responses_from_details_date_id,
     )
 
     # TODO (nathan): better handling of model_args

src/lighteval/main_endpoint.py

@@ -179,6 +179,9 @@ def inference_endpoint(
     num_fewshot_seeds: Annotated[
         int, Option(help="Number of seeds to use for few-shot evaluation.", rich_help_panel=HELP_PANEL_NAME_1)
     ] = 1,
+    load_responses_from_details_date_id: Annotated[
+        Optional[str], Option(help="Load responses from details directory.", rich_help_panel=HELP_PANEL_NAME_1)
+    ] = None,
     # === saving ===
     output_dir: Annotated[
         str, Option(help="Output directory for evaluation results.", rich_help_panel=HELP_PANEL_NAME_2)
@@ -247,6 +250,7 @@ def inference_endpoint(
         max_samples=max_samples,
         use_chat_template=use_chat_template,
         system_prompt=system_prompt,
+        load_responses_from_details_date_id=load_responses_from_details_date_id,
     )
     pipeline = Pipeline(
         tasks=tasks,
@@ -292,6 +296,9 @@ def tgi(
     num_fewshot_seeds: Annotated[
         int, Option(help="Number of seeds to use for few-shot evaluation.", rich_help_panel=HELP_PANEL_NAME_1)
     ] = 1,
+    load_responses_from_details_date_id: Annotated[
+        Optional[str], Option(help="Load responses from details directory.", rich_help_panel=HELP_PANEL_NAME_1)
+    ] = None,
     # === saving ===
     output_dir: Annotated[
         str, Option(help="Output directory for evaluation results.", rich_help_panel=HELP_PANEL_NAME_2)
@@ -355,6 +362,7 @@ def tgi(
         max_samples=max_samples,
         use_chat_template=use_chat_template,
         system_prompt=system_prompt,
+        load_responses_from_details_date_id=load_responses_from_details_date_id,
     )
     pipeline = Pipeline(
         tasks=tasks,
@@ -400,6 +408,9 @@ def litellm(
     num_fewshot_seeds: Annotated[
         int, Option(help="Number of seeds to use for few-shot evaluation.", rich_help_panel=HELP_PANEL_NAME_1)
     ] = 1,
+    load_responses_from_details_date_id: Annotated[
+        Optional[str], Option(help="Load responses from details directory.", rich_help_panel=HELP_PANEL_NAME_1)
+    ] = None,
     # === saving ===
     output_dir: Annotated[
         str, Option(help="Output directory for evaluation results.", rich_help_panel=HELP_PANEL_NAME_2)
@@ -464,6 +475,7 @@ def litellm(
         max_samples=max_samples,
         use_chat_template=use_chat_template,
         system_prompt=system_prompt,
+        load_responses_from_details_date_id=load_responses_from_details_date_id,
     )
     pipeline = Pipeline(
         tasks=tasks,

src/lighteval/main_vllm.py

@@ -63,6 +63,9 @@ def vllm(
     num_fewshot_seeds: Annotated[
         int, Option(help="Number of seeds to use for few-shot evaluation.", rich_help_panel=HELP_PANEL_NAME_1)
     ] = 1,
+    load_responses_from_details_date_id: Annotated[
+        Optional[str], Option(help="Load responses from details directory.", rich_help_panel=HELP_PANEL_NAME_1)
+    ] = None,
     # === saving ===
     output_dir: Annotated[
         str, Option(help="Output directory for evaluation results.", rich_help_panel=HELP_PANEL_NAME_2)
@@ -124,6 +127,7 @@ def vllm(
         max_samples=max_samples,
         use_chat_template=use_chat_template,
         system_prompt=system_prompt,
+        load_responses_from_details_date_id=load_responses_from_details_date_id,
     )
 
     if model_args.endswith(".yaml"):

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/imports/__init__.py

@@ -0,0 +1,21 @@
+# MIT License
+
+# Copyright (c) 2024 The HuggingFace Team
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.

src/lighteval/metrics/utils/__init__.py

@@ -0,0 +1,21 @@
+# MIT License
+
+# Copyright (c) 2024 The HuggingFace Team
+
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.

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: