Adds RULER benchmark

src/lighteval/metrics/metrics.py

@@ -133,7 +133,27 @@ class Metrics(Enum):
         corpus_level_fn=np.mean,
         higher_is_better=True,
     )
-
+    ruler_match_any = SampleLevelMetric(
+        metric_name="ruler_match_any",
+        sample_level_fn=lambda predictions, golds, formatted_doc: max(
+            [1.0 if r.lower() in predictions[0].lower() else 0.0 for r in golds]
+        ),
+        category=MetricCategory.GENERATIVE,
+        use_case=MetricUseCase.SUMMARIZATION,
+        corpus_level_fn=np.mean,
+        higher_is_better=True,
+    )
+    ruler_match_all = SampleLevelMetric(
+        metric_name="ruler_match_all",
+        sample_level_fn=lambda predictions, golds, formatted_doc: sum(
+            [1.0 if r.lower() in predictions[0].lower() else 0.0 for r in golds]
+        )
+        / len(golds),
+        category=MetricCategory.GENERATIVE,
+        use_case=MetricUseCase.SUMMARIZATION,
+        corpus_level_fn=np.mean,
+        higher_is_better=True,
+    )
     bleurt = SampleLevelMetric(
         metric_name="bleurt",
         sample_level_fn=BLEURT().compute,

src/lighteval/models/vllm/vllm_model.py

@@ -105,6 +105,8 @@ class VLLMModelConfig(ModelConfig):
     max_num_batched_tokens: PositiveInt = 2048  # maximum number of tokens per batch
     subfolder: str | None = None
     is_async: bool = False  # Whether to use the async version or sync version of the model
+    use_dual_chunk_attention: bool = False
+    enforce_eager: bool = False
 
 
 class VLLMModel(LightevalModel):
@@ -187,6 +189,8 @@ def _create_auto_model(self, config: VLLMModelConfig) -> Optional[LLM]:
             "seed": int(config.seed),
             "max_num_seqs": int(config.max_num_seqs),
             "max_num_batched_tokens": int(config.max_num_batched_tokens),
+            "enforce_eager": config.enforce_eager,
+            "use_dual_chunk_attention": config.use_dual_chunk_attention,
         }
 
         if config.quantization is not None:
@@ -276,7 +280,7 @@ def greedy_until(
             if max_new_tokens is not None:
                 if context_size + max_new_tokens > self.max_length:
                     logger.warning(
-                        f"{context_size + max_new_tokens=} which is greater than {self.max_length=}. Truncating context to {self.max_length - max_new_tokens} tokens."
+                        f"{context_size + max_new_tokens=} which is greater than {self.max_length=}. Truncating context to {self.max_length=} - {max_new_tokens=} = {self.max_length - max_new_tokens} tokens."
                     )
                     context_size = self.max_length - max_new_tokens
                     if context_size < 0:

src/lighteval/tasks/default_prompts.py

@@ -44,6 +44,15 @@
 # fmt: on
 
 
+def ruler(line, task_name: str = None):
+    query = line["input"]
+    choices = line["outputs"]
+    gold_index = 0
+    instruction = "Only answer the question to complete the prompt, without any additional text.\n"
+    query = f"{instruction}{query}"
+
+    return Doc(query=query, instruction=instruction, choices=choices, gold_index=gold_index, task_name=task_name)
+
 def mmmu_pro(line, task_name: Optional[str] = None):
     # fmt: off
     question = line["question"]        # "What is the capital of France?"
@@ -87,7 +96,6 @@ def mmmu_pro(line, task_name: Optional[str] = None):
         instruction=instructions,
     )
 
-
 def mmmu_pro_vision(line, task_name: str = None):
     instruction = (
         "Answer with the option letter from the given choices directly."
@@ -119,14 +127,17 @@ def mmmu_pro_vision(line, task_name: str = None):
         instruction=instruction,
     )
 
-
 def simpleqa(line, task_name: str = None):
     query = line["problem"]
     choices = [line["answer"]]
     gold_index = 0
 
     return Doc(
-        task_name=task_name, query=query, choices=choices, gold_index=gold_index, specific={**eval(line["metadata"])}
+        task_name=task_name,
+        query=query,
+        choices=choices,
+        gold_index=gold_index,
+        specific={**eval(line["metadata"])},
     )
 
 

src/lighteval/tasks/extended/__init__.py

@@ -30,9 +30,10 @@
     import lighteval.tasks.extended.mix_eval.main as mix_eval
     import lighteval.tasks.extended.mt_bench.main as mt_bench
     import lighteval.tasks.extended.olympiade_bench.main as olympiad_bench
+    import lighteval.tasks.extended.ruler.main as ruler
     import lighteval.tasks.extended.tiny_benchmarks.main as tiny_benchmarks
 
-    AVAILABLE_EXTENDED_TASKS_MODULES = [ifeval, tiny_benchmarks, mt_bench, mix_eval, olympiad_bench, hle, lcb]
+    AVAILABLE_EXTENDED_TASKS_MODULES = [ifeval, tiny_benchmarks, mt_bench, mix_eval, olympiad_bench, hle, lcb, ruler]
 
 else:
     AVAILABLE_EXTENDED_TASKS_MODULES = []

src/lighteval/tasks/extended/ruler/main.py

@@ -0,0 +1,69 @@
+# 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.
+
+import lighteval.tasks.default_prompts as prompt
+from lighteval.metrics.metrics import Metrics
+from lighteval.tasks.lighteval_task import LightevalTaskConfig
+
+
+subsets = [
+    "niah_single_1",
+    "niah_single_2",
+    "niah_single_3",
+    "niah_multikey_1",
+    "niah_multikey_2",
+    "niah_multikey_3",
+    "niah_multiquery",
+    "niah_multivalue",
+    "vt",
+    "cwe",
+    "fwe",
+    "qa_1",
+    "qa_2",
+]
+
+lengths = [131072, 65536, 32768, 16384, 8192, 4096]
+
+task_configs = []
+
+for subset in subsets:
+    for length in lengths:
+        task_configs.append(
+            LightevalTaskConfig(
+                name=f"ruler_{length}:{subset}",
+                suite=["lighteval"],
+                prompt_function=prompt.ruler,
+                hf_repo=f"SaylorTwift/RULER-{length}-llama-3.2-tokenizer",
+                hf_subset="default",
+                hf_avail_splits=[subset],
+                evaluation_splits=[subset],
+                few_shots_split=None,
+                few_shots_select=None,
+                generation_size=128 if "niah" in subset else 30 if subset == "vt" else 120 if subset == "cwe" else 50,
+                metric=[Metrics.ruler_match_any] if subset in ["qa_1", "qa_2"] else [Metrics.ruler_match_all],
+                stop_sequence=None,
+                trust_dataset=False,
+                version=0,
+            )
+        )
+
+TASKS_TABLE = task_configs

src/lighteval/tasks/lighteval_task.py

@@ -31,6 +31,7 @@
 from huggingface_hub import TextGenerationInputGrammarType
 from multiprocess import Pool
 from pytablewriter import MarkdownTableWriter
+from tqdm import tqdm
 
 from lighteval.metrics import (
     apply_generative_metric,
@@ -560,7 +561,7 @@ def load_datasets(tasks: list["LightevalTask"], dataset_loading_processes: int =
                     task.dataset_filter,
                     task.dataset_revision,
                 )
-                for task in tasks
+                for task in tqdm(tasks)
             ]
         else:
             with Pool(processes=dataset_loading_processes) as pool:
@@ -627,7 +628,7 @@ def create_requests_from_tasks(  # noqa: C901
     task_dict_items = [(name, task) for name, task in task_dict.items() if len(task.eval_docs()) > 0]
 
     # Get lists of each type of request
-    for task_name, task in task_dict_items:
+    for task_name, task in tqdm(task_dict_items):
         task_docs = list(task.eval_docs())
         n_samples = min(max_samples, len(task_docs)) if max_samples else len(task_docs)
         evaluation_tracker.task_config_logger.log_num_docs(task_name, len(task_docs), n_samples)