Add --examples Argument for Fine-Grained Task Evaluation in lm-evaluation-harness. This feature is the first step towards efficient multi-prompt evaluation with PromptEval [1,2]

lm_eval/__main__.py

@@ -128,6 +128,15 @@ def setup_parser() -> argparse.ArgumentParser:
         help="Limit the number of examples per task. "
         "If <1, limit is a percentage of the total number of examples.",
     )
+    parser.add_argument(
+        "--examples",
+        "-E",
+        default=None,
+        type=str,
+        metavar="/path/to/json",
+        help="Examples to test. "
+        "Should be a json file which loads into a Python dictionary. E.g., {'mmlu_anatomy':[0,1],'mmlu_astronomy':[1,2,3]}.",
+    )
     parser.add_argument(
         "--use_cache",
         "-c",
@@ -314,10 +323,18 @@ def cli_evaluate(args: Union[argparse.Namespace, None] = None) -> None:
         )
 
     if args.limit:
+        limit = args.limit
         eval_logger.warning(
             " --limit SHOULD ONLY BE USED FOR TESTING."
             "REAL METRICS SHOULD NOT BE COMPUTED USING LIMIT."
         )
+    if args.examples:
+        assert (
+            args.limit is None
+        ), "If --examples is not None, then --limit must be None."
+        limit = None
+        with open(args.examples, "r") as json_file:
+            examples = json.load(json_file)
 
     if args.tasks is None:
         eval_logger.error("Need to specify task to evaluate.")
@@ -393,7 +410,8 @@ def cli_evaluate(args: Union[argparse.Namespace, None] = None) -> None:
         max_batch_size=args.max_batch_size,
         device=args.device,
         use_cache=args.use_cache,
-        limit=args.limit,
+        limit=limit,
+        examples=examples,
         check_integrity=args.check_integrity,
         write_out=args.write_out,
         log_samples=args.log_samples,
@@ -451,7 +469,7 @@ def cli_evaluate(args: Union[argparse.Namespace, None] = None) -> None:
             evaluation_tracker.recreate_metadata_card()
 
         print(
-            f"{args.model} ({args.model_args}), gen_kwargs: ({args.gen_kwargs}), limit: {args.limit}, num_fewshot: {args.num_fewshot}, "
+            f"{args.model} ({args.model_args}), gen_kwargs: ({args.gen_kwargs}), limit: {limit}, num_fewshot: {args.num_fewshot}, "
             f"batch_size: {args.batch_size}{f' ({batch_sizes})' if batch_sizes else ''}"
         )
         print(make_table(results))

lm_eval/api/task.py

@@ -374,6 +374,7 @@ def build_all_requests(
         self,
         *,
         limit: Union[int, None] = None,
+        examples: Optional[List[int]] = None,
         rank: int = 0,
         world_size: int = 1,
         cache_requests: bool = False,
@@ -426,7 +427,9 @@ def build_all_requests(
             limit = None
 
         doc_id_docs = list(
-            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
+            self.doc_iterator(
+                rank=rank, limit=limit, examples=examples, world_size=world_size
+            )
         )
 
         num_docs = len(doc_id_docs)
@@ -678,15 +681,38 @@ def eval_docs(self) -> Union[datasets.Dataset, List[dict]]:
             )
 
     def doc_iterator(
-        self, *, rank: int = 0, limit: Union[int, None] = None, world_size: int = 1
+        self,
+        *,
+        rank: int = 0,
+        limit: Union[int, None] = None,
+        examples: Optional[List[int]] = None,
+        world_size: int = 1,
     ) -> Iterator[Tuple[int, Any]]:
-        limit = int(limit) if limit else None
-        doc_iterator = utils.create_iterator(
-            enumerate(self.eval_docs),
-            rank=int(rank),
-            limit=limit,
-            world_size=int(world_size),
-        )
+        if examples:
+            n = self.eval_docs.to_pandas().shape[0]
+            assert all(
+                [e < n for e in examples]
+            ), f"Elements of --examples should be in the interval [0,k-1] where k is the number of total examples. In this case, k={n}."
+            doc_iterator = utils.create_iterator(
+                enumerate(
+                    datasets.Dataset.from_pandas(
+                        self.eval_docs.to_pandas()
+                        .iloc[examples, :]
+                        .reset_index(drop=True)
+                    )
+                ),
+                rank=int(rank),
+                limit=None,  # limit does not matter here since we are selecting samples directly
+                world_size=int(world_size),
+            )
+        else:
+            limit = int(limit) if limit else None
+            doc_iterator = utils.create_iterator(
+                enumerate(self.eval_docs),
+                rank=int(rank),
+                limit=limit,
+                world_size=int(world_size),
+            )
         return doc_iterator
 
 

lm_eval/evaluator.py

@@ -26,10 +26,7 @@
 )
 from lm_eval.loggers import EvaluationTracker
 from lm_eval.loggers.utils import add_env_info, add_tokenizer_info, get_git_commit_hash
-from lm_eval.tasks import (
-    TaskManager,
-    get_task_dict,
-)
+from lm_eval.tasks import TaskManager, get_task_dict
 from lm_eval.utils import (
     eval_logger,
     handle_non_serializable,
@@ -58,6 +55,7 @@ def simple_evaluate(
     rewrite_requests_cache: bool = False,
     delete_requests_cache: bool = False,
     limit: Optional[Union[int, float]] = None,
+    examples: Optional[dict] = None,
     bootstrap_iters: int = 100000,
     check_integrity: bool = False,
     write_out: bool = False,
@@ -103,6 +101,8 @@ def simple_evaluate(
         Deletes all of the request cache if set to `True`. `None` if not desired.
     :param limit: int or float, optional
         Limit the number of examples per task (only use this for testing), If <1, limit is a percentage of the total number of examples.
+    :param examples: dictionary, optional
+        Dictionary indicating which examples should be tested in each task, e.g., {'mmlu_astronomy':[0,3,6],'mmlu_anatomy':[1,4,7,10]}.
     :param bootstrap_iters:
         Number of iterations for bootstrap statistics, used when calculating stderrs. set to 0 for no stderr calculations to be performed.
     :param check_integrity: bool
@@ -140,6 +140,11 @@ def simple_evaluate(
     eval_logger.setLevel(getattr(logging, f"{verbosity}"))
     start_date = time.time()
 
+    if limit is not None and examples is not None:
+        raise ValueError(
+            "Either 'limit' or 'examples' must be None, but both are not None."
+        )
+
     if delete_requests_cache:
         eval_logger.info("Deleting requests cache...")
         delete_cache()
@@ -305,6 +310,7 @@ def _adjust_config(task_dict):
         lm=lm,
         task_dict=task_dict,
         limit=limit,
+        examples=examples,
         cache_requests=cache_requests,
         rewrite_requests_cache=rewrite_requests_cache,
         bootstrap_iters=bootstrap_iters,
@@ -365,6 +371,7 @@ def evaluate(
     lm: "LM",
     task_dict,
     limit: Optional[int] = None,
+    examples: Optional[dict] = None,
     cache_requests: bool = False,
     rewrite_requests_cache: bool = False,
     bootstrap_iters: Optional[int] = 100000,
@@ -384,6 +391,8 @@ def evaluate(
         Dictionary of tasks. Tasks will be taken to have name type(task).config.task .
     :param limit: int, optional
         Limit the number of examples per task (only use this for testing)
+    :param examples: dictionary, optional
+        Dictionary indicating which examples should be tested in each task, e.g., {'mmlu_astronomy':[0,3,6],'mmlu_anatomy':[1,4,7,10]}.
     :param cache_requests: bool, optional
         Speed up evaluation by caching the building of dataset requests.
     :param rewrite_requests_cache: bool, optional
@@ -413,11 +422,14 @@ def evaluate(
 
     eval_logger.setLevel(getattr(logging, f"{verbosity}"))
 
+    if limit is not None and examples is not None:
+        raise ValueError(
+            "Either 'limit' or 'examples' must be None, but both are not None."
+        )
     if apply_chat_template:
         eval_logger.warning(
             "Chat template formatting change affects loglikelihood and multiple-choice tasks. See docs/chat-template-readme.md for details."
         )
-
     # tracks all Instances/requests a model must generate output on.
     requests = defaultdict(list)
     # stores the amount to pad out reqs per req. type so that
@@ -467,6 +479,9 @@ def evaluate(
         limits.append(limit)
         task.build_all_requests(
             limit=limit,
+            examples=examples[task_output.task_name]
+            if examples is not None
+            else examples,
             rank=lm.rank,
             world_size=lm.world_size,
             cache_requests=cache_requests,
@@ -551,17 +566,26 @@ def evaluate(
         # iterate over different filters used
         for filter_key in task.instances[0].filtered_resps.keys():
             doc_iterator = task.doc_iterator(
-                rank=RANK, limit=limit, world_size=WORLD_SIZE
+                rank=RANK,
+                limit=limit,
+                examples=examples[task_output.task_name]
+                if examples is not None
+                else examples,
+                world_size=WORLD_SIZE,
             )
             for doc_id, doc in doc_iterator:
+                if examples:
+                    doc_id_true = examples[task_output.task_name][doc_id]
+                else:
+                    doc_id_true = doc_id
                 requests = instances_by_doc_id[doc_id]
                 metrics = task.process_results(
                     doc, [req.filtered_resps[filter_key] for req in requests]
                 )
                 if log_samples:
                     target = task.doc_to_target(doc)
                     example = {
-                        "doc_id": doc_id,
+                        "doc_id": doc_id_true,
                         "doc": doc,
                         "target": target,
                         "arguments": [req.args for req in requests],