ai2-olmo-eval 0.7.2__py3-none-any.whl → 0.8.0__py3-none-any.whl

{ai2_olmo_eval-0.7.2.dist-info → ai2_olmo_eval-0.8.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ai2-olmo-eval
-Version: 0.7.2
+Version: 0.8.0
 Summary: In-loop evaluation tasks for language modeling
 Author-email: Allen Institute for Artificial Intelligence <olmo@allenai.org>
 License:                                  Apache License

{ai2_olmo_eval-0.7.2.dist-info → ai2_olmo_eval-0.8.0.dist-info}/RECORD

@@ -1,10 +1,10 @@
-ai2_olmo_eval-0.7.2.dist-info/licenses/LICENSE,sha256=YvuKOpYh3COIF0yqq-nCMXtpS7mh1GyYvPVlW2j1G-M,11359
+ai2_olmo_eval-0.8.0.dist-info/licenses/LICENSE,sha256=YvuKOpYh3COIF0yqq-nCMXtpS7mh1GyYvPVlW2j1G-M,11359
 olmo_eval/__init__.py,sha256=49RxnAaJNk8U9XP3SF5MjyFIxLSkxH0vXQuZgnEOi44,283
-olmo_eval/metrics.py,sha256=NcI_1B3BV-DC9RXjsSIftU-2GeF8vvU6SNyJnlYlKwU,18705
-olmo_eval/tasks.py,sha256=QGLyF7JA2-T9mkh-N4cZGNOQp9si90yQSS41T3x5Lak,79630
+olmo_eval/metrics.py,sha256=zc4IOZ8rUhxPyXVk6fOYzVKjJ4Lzq4tYeoyurxYWqY0,20034
+olmo_eval/tasks.py,sha256=DF4-2MS5dkGgZSjNrRkjEoWShrAsGO7tiB6mqcTQnE8,93219
 olmo_eval/tokenizer.py,sha256=PnkidE0nAtEA1QZjuQpE_bIwgAsHxodnaJRALAPqrJQ,5127
 olmo_eval/util.py,sha256=ARmZmRQl8VOvnKQoUprb3cOunzcApeNhRdV4BMXZuvo,3856
-olmo_eval/version.py,sha256=QWjPfx79C2NOQw2G7iDEsM4FKsLiGLCLNDzEx7EImf8,308
+olmo_eval/version.py,sha256=ucNFr1ahYQCmPHuM8Qq6kPbT7lmTnsZQuSxG1jpqphI,308
 olmo_eval/hf_datasets/ai2_arc/ARC-Challenge/validation/data-00000-of-00001.arrow,sha256=TPWbMhBmticWjYp7TA3etcKbXbaoCDBWhxuqlD1bDJA,98080
 olmo_eval/hf_datasets/ai2_arc/ARC-Challenge/validation/dataset_info.json,sha256=iZumP5Udu8LD7cbew3o7nNpnGu-o9jPaMxUrNDDNIVY,1795
 olmo_eval/hf_datasets/ai2_arc/ARC-Challenge/validation/state.json,sha256=6Q1XhM-HMZcymuGAKBC_8RjMBKgJSaR_6lLUO9Z8XwE,255
@@ -716,7 +716,7 @@ olmo_eval/oe_eval_tasks/winogrande/val_rc_5shot/config.json,sha256=ySjEVqTOj5GwC
 olmo_eval/oe_eval_tasks/winogrande/val_rc_5shot/requests.jsonl.gz,sha256=knTzcqigWCfdYLN1Pl0TfCm0Fi1lRASWAo_SC6KtXsc,115262
 olmo_eval/tokenizers/allenai_eleuther-ai-gpt-neox-20b-pii-special.json,sha256=yjXYcnpTO7Zjm_R4Gucrn9oA5paadiYM-ZZER5q_EXc,2114319
 olmo_eval/tokenizers/allenai_gpt-neox-olmo-dolma-v1_5.json,sha256=mtM7Szmp-Dlzw_jEKgGUjdW4d6KKyaU1aVbE_07QtxQ,2115113
-ai2_olmo_eval-0.7.2.dist-info/METADATA,sha256=PKJfkoDu4hrLzb6NA1MDfXOjZnUxQ4WFpJouWU1Cr_4,14398
-ai2_olmo_eval-0.7.2.dist-info/WHEEL,sha256=Nw36Djuh_5VDukK0H78QzOX-_FQEo6V37m3nkm96gtU,91
-ai2_olmo_eval-0.7.2.dist-info/top_level.txt,sha256=Pryk28JTb89-j624Uy1gRZiE0YXI3czgbNIfJCl9-x0,10
-ai2_olmo_eval-0.7.2.dist-info/RECORD,,
+ai2_olmo_eval-0.8.0.dist-info/METADATA,sha256=TZmOipbol7scpsNfiFVximYmOONNlOg-J_bhbn0a-FE,14398
+ai2_olmo_eval-0.8.0.dist-info/WHEEL,sha256=Nw36Djuh_5VDukK0H78QzOX-_FQEo6V37m3nkm96gtU,91
+ai2_olmo_eval-0.8.0.dist-info/top_level.txt,sha256=Pryk28JTb89-j624Uy1gRZiE0YXI3czgbNIfJCl9-x0,10
+ai2_olmo_eval-0.8.0.dist-info/RECORD,,

olmo_eval/metrics.py

@@ -98,96 +98,121 @@ class ICLMetric(Metric):
                 batch["ctx_len"][idx] - 1 : batch["ctx_len"][idx] + batch["cont_len"][idx] - 1
             ]
 
-            log_likelihood: torch.Tensor
-            celoss: torch.Tensor
-            bpb: torch.Tensor
-            log_likelihood_no_leading_space: torch.Tensor
-            celoss_no_leading_space: torch.Tensor
-            bpb_no_leading_space: torch.Tensor
-            if self.metric_type == "pmi_dc":
-                assert dc_lm_logits is not None
-                # get domain conditional continuation logits: [cont_len, vocab]
-                dc_lm_cont_logits = dc_lm_logits[idx][
-                    batch["dc_len"][idx] - 1 : batch["dc_len"][idx] + batch["cont_len"][idx] - 1
-                ]
-
-                # gather log-probs at continuation token indices but divide by domain conditional prob
-                log_likelihood = (
-                    torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / torch.gather(dc_lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                )
-                celoss = -log_likelihood
-                bpb = -log_likelihood  # the normalization factors cancel out
-
-                log_likelihood_no_leading_space = log_likelihood
-                celoss_no_leading_space = celoss
-                bpb_no_leading_space = bpb
-            elif self.metric_type == "acc" or self.metric_type == "f1":
-                # gather log-probs at continuation token indices
-                log_likelihood = torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                celoss = (
-                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / batch["cont_str_len"][idx]
-                )
-                bpb = (
-                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / batch["cont_byte_len"][idx]
-                    * LOG_2_OF_E
-                )
-
-                log_likelihood_no_leading_space = torch.gather(
-                    lm_cont_logits, 1, cont_tokens.unsqueeze(-1)
-                ).sum()
-                celoss_no_leading_space = (
-                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / batch["cont_str_len_no_leading_space"][idx]
-                )
-                bpb_no_leading_space = (
-                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / batch["cont_byte_len_no_leading_space"][idx]
-                    * LOG_2_OF_E
-                )
-            elif self.metric_type in ["len_norm", "ce_loss", "bpb"]:
-                log_likelihood = (
-                    torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / batch["cont_str_len"][idx]
-                )
-                celoss = (
-                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / batch["cont_str_len"][idx]
-                )
-                bpb = (
-                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / batch["cont_byte_len"][idx]
-                    * LOG_2_OF_E
-                )
+            if "choice_ids" in batch:
+                fast_mc = True
+                choice_ids = batch["choice_ids"][idx]
+            else:
+                fast_mc = False
+                choice_ids = cont_tokens
+
+            # For each choice token, calculate metrics and append as separate entries
+            for choice_idx, choice_token in enumerate(choice_ids):
+                if fast_mc:
+                    _cont_id = choice_idx
+                    _cont_tokens = choice_token.unsqueeze(-1)
+                else:
+                    _cont_id = cont_id
+                    _cont_tokens = cont_tokens
+
+                # Skip choices for Qs with less than the max choices (for questions w/ different nubmers of choices)
+                is_empty_choice = (choice_token.unsqueeze(-1).unsqueeze(-1) == -1).all().item()
+                if is_empty_choice:
+                    continue
+
+                log_likelihood: torch.Tensor
+                celoss: torch.Tensor
+                bpb: torch.Tensor
+                log_likelihood_no_leading_space: torch.Tensor
+                celoss_no_leading_space: torch.Tensor
+                bpb_no_leading_space: torch.Tensor
+                if self.metric_type == "pmi_dc":
+                    assert dc_lm_logits is not None
+                    # get domain conditional continuation logits: [cont_len, vocab]
+                    dc_lm_cont_logits = dc_lm_logits[idx][
+                        batch["dc_len"][idx] - 1 : batch["dc_len"][idx] + batch["cont_len"][idx] - 1
+                    ]
 
-                log_likelihood_no_leading_space = (
-                    torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / batch["cont_str_len_no_leading_space"][idx]
-                )
-                celoss_no_leading_space = (
-                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / batch["cont_str_len_no_leading_space"][idx]
+                    # gather log-probs at continuation token indices but divide by domain conditional prob
+                    log_likelihood = (
+                        torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / torch.gather(dc_lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                    )
+                    celoss = -log_likelihood
+                    bpb = -log_likelihood  # the normalization factors cancel out
+
+                    log_likelihood_no_leading_space = log_likelihood
+                    celoss_no_leading_space = celoss
+                    bpb_no_leading_space = bpb
+                elif self.metric_type == "acc" or self.metric_type == "f1":
+                    # gather log-probs at continuation token indices
+                    log_likelihood = torch.gather(
+                        lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)
+                    ).sum()
+                    celoss = (
+                        -torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / batch["cont_str_len"][idx]
+                    )
+                    bpb = (
+                        -torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / batch["cont_byte_len"][idx]
+                        * LOG_2_OF_E
+                    )
+
+                    log_likelihood_no_leading_space = torch.gather(
+                        lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)
+                    ).sum()
+                    celoss_no_leading_space = (
+                        -torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / batch["cont_str_len_no_leading_space"][idx]
+                    )
+                    bpb_no_leading_space = (
+                        -torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / batch["cont_byte_len_no_leading_space"][idx]
+                        * LOG_2_OF_E
+                    )
+                elif self.metric_type in ["len_norm", "ce_loss", "bpb"]:
+                    log_likelihood = (
+                        torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / batch["cont_str_len"][idx]
+                    )
+                    celoss = (
+                        -torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / batch["cont_str_len"][idx]
+                    )
+                    bpb = (
+                        -torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / batch["cont_byte_len"][idx]
+                        * LOG_2_OF_E
+                    )
+
+                    log_likelihood_no_leading_space = (
+                        torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / batch["cont_str_len_no_leading_space"][idx]
+                    )
+                    celoss_no_leading_space = (
+                        -torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / batch["cont_str_len_no_leading_space"][idx]
+                    )
+                    bpb_no_leading_space = (
+                        -torch.gather(lm_cont_logits, 1, _cont_tokens.unsqueeze(-1)).sum()
+                        / batch["cont_byte_len_no_leading_space"][idx]
+                        * LOG_2_OF_E
+                    )
+                else:
+                    raise ValueError(self.metric_type)
+
+                self.labels.append((doc_id, _cont_id, int(batch["label_id"][idx])))
+                self.loglikelihoods.append((doc_id, _cont_id, float(log_likelihood)))
+                self.celosses.append((doc_id, _cont_id, float(celoss)))
+                self.bpbs.append((doc_id, _cont_id, float(bpb)))
+
+                self.loglikelihoods_no_leading_space.append(
+                    (doc_id, _cont_id, float(log_likelihood_no_leading_space))
                 )
-                bpb_no_leading_space = (
-                    -torch.gather(lm_cont_logits, 1, cont_tokens.unsqueeze(-1)).sum()
-                    / batch["cont_byte_len_no_leading_space"][idx]
-                    * LOG_2_OF_E
+                self.celosses_no_leading_space.append(
+                    (doc_id, _cont_id, float(celoss_no_leading_space))
                 )
-            else:
-                raise ValueError(self.metric_type)
-
-            self.labels.append((doc_id, cont_id, int(batch["label_id"][idx])))
-            self.loglikelihoods.append((doc_id, cont_id, float(log_likelihood)))
-            self.celosses.append((doc_id, cont_id, float(celoss)))
-            self.bpbs.append((doc_id, cont_id, float(bpb)))
-
-            self.loglikelihoods_no_leading_space.append(
-                (doc_id, cont_id, float(log_likelihood_no_leading_space))
-            )
-            self.celosses_no_leading_space.append((doc_id, cont_id, float(celoss_no_leading_space)))
-            self.bpbs_no_leading_space.append((doc_id, cont_id, float(bpb_no_leading_space)))
+                self.bpbs_no_leading_space.append((doc_id, _cont_id, float(bpb_no_leading_space)))
 
     def compute(self) -> Dict[str, torch.Tensor]:
         # Task "suffix" -> tensor

olmo_eval/tasks.py

@@ -33,6 +33,7 @@ class ICLMultiChoiceTaskDataset(metaclass=abc.ABCMeta):
         dataset_name: Union[str, Sequence[str], None] = None,
         model_ctx_len: int = 2048,
         fixed_ctx_len: bool = False,
+        fast_mc: bool = False,
         split="validation",
         metric_type=None,  # Override default metric type
         prompts: Optional[List[Optional[str]]] = None,  # List of prompt variants to use
@@ -44,6 +45,7 @@ class ICLMultiChoiceTaskDataset(metaclass=abc.ABCMeta):
         self.dataset_name = dataset_name
         self.model_ctx_len = model_ctx_len
         self.fixed_ctx_len = fixed_ctx_len
+        self.fast_mc = fast_mc
         self.prompts = prompts or [None]
         self.current_prompt: Optional[str] = None
         if metric_type is not None:
@@ -76,6 +78,7 @@ class ICLMultiChoiceTaskDataset(metaclass=abc.ABCMeta):
     def prep_examples(self):
         """Append doc_ids to each example so that they are processed together in the metric"""
         doc_id = 0
+        new_samples = []
         for doc in self.dataset:
             for prompt in self.prompts:
                 self.current_prompt = prompt
@@ -125,7 +128,7 @@ class ICLMultiChoiceTaskDataset(metaclass=abc.ABCMeta):
                     dc_query = dc + continuation[:-1]
 
                     # form a sample
-                    self.samples.append(
+                    new_samples.append(
                         {
                             "doc_id": doc_id,
                             "cont_id": cont_id,
@@ -148,6 +151,56 @@ class ICLMultiChoiceTaskDataset(metaclass=abc.ABCMeta):
 
                 doc_id += 1
 
+        # Fast MCQA:
+        # Only pass a single request, and group together all continuations as tokens
+        if self.fast_mc:
+            # Get unique doc IDs
+            unique_doc_ids = {
+                sample["doc_id"] for sample in new_samples if isinstance(sample["doc_id"], int)
+            }
+
+            # Create new samples list for fast MC
+            fast_mc_samples = []
+
+            # Process each unique document
+            for doc_id in unique_doc_ids:
+                # Get all samples for this doc_id
+                doc_samples = [s for s in new_samples if s["doc_id"] == doc_id]
+
+                # Sort by continuation ID
+                doc_samples.sort(
+                    key=lambda x: float(x["cont_id"])
+                    if isinstance(x["cont_id"], (int, float))
+                    else 0.0
+                )
+
+                # Create new sample with distractor continuations
+                base_sample = doc_samples[0].copy()
+                choices = [s["continuation"] for s in doc_samples]
+
+                # Assert all continuations are length 1
+                for choice in choices:
+                    if not isinstance(choice, (list, tuple)):
+                        raise TypeError(
+                            f"Expected continuation to be a list or tuple, got {type(choice)}"
+                        )
+                    assert len(choice) == 1, f"Expected continuation length 1, got {len(choice)}"
+
+                # Take first token of each continuation
+                choices = [
+                    choice[0] if isinstance(choice, (list, tuple)) else choice for choice in choices
+                ]
+
+                base_sample["choices"] = choices
+                base_sample["fast_mc"] = True
+
+                fast_mc_samples.append(base_sample)
+
+            # Add fast MC samples to main samples list
+            new_samples = fast_mc_samples
+
+        self.samples = new_samples
+
     def pad_tokens_until_max(self, tokens, max_len=2048):
         """truncate from left if len(tokens) > model_ctx_len, max_len is not considered then
         queries are already truncated at max length of model_ctx_len
@@ -214,6 +267,7 @@ class ICLMultiChoiceTaskDataset(metaclass=abc.ABCMeta):
         ctxs = []
         continuations = []
         ctx_lens = []
+        choice_ids = []
         dc_lens = []
         cont_lens = []
         cont_str_lens = []
@@ -245,6 +299,8 @@ class ICLMultiChoiceTaskDataset(metaclass=abc.ABCMeta):
             cont_byte_lens.append(sample["cont_byte_len"])
             cont_str_len_no_leading_space.append(sample["cont_str_len_no_leading_space"])
             cont_byte_len_no_leading_space.append(sample["cont_byte_len_no_leading_space"])
+            if self.fast_mc:
+                choice_ids.append(sample["choices"])
 
             queries.append(
                 torch.LongTensor(
@@ -281,6 +337,16 @@ class ICLMultiChoiceTaskDataset(metaclass=abc.ABCMeta):
             "label_id": torch.LongTensor(label_ids),
         }
 
+        if self.fast_mc:
+            # Pad choice_ids with -1 (for Qs with different numbers of choices)
+            max_choices_len = max(len(choices) for choices in choice_ids)
+            padded_choice_ids = []
+            for choices in choice_ids:
+                padding = [-1] * (max_choices_len - len(choices))
+                padded_choice_ids.append(choices + padding)
+            choice_ids = padded_choice_ids
+            batch["choice_ids"] = torch.LongTensor(choice_ids)
+
         return batch
 
     def token_encode(self, string: str) -> List[int]:
@@ -1446,6 +1512,7 @@ class OEEvalTask(ICLMultiChoiceTaskDataset):
         dataset_name: Union[str, Sequence[str], None] = None,
         model_ctx_len: int = 2048,
         fixed_ctx_len: bool = False,
+        fast_mc: bool = False,
         split=None,
         metric_type=None,
         prompts: Optional[List[Optional[str]]] = None,  # List of prompt variants to use
@@ -1457,6 +1524,7 @@ class OEEvalTask(ICLMultiChoiceTaskDataset):
         self.dataset_name = dataset_name
         self.model_ctx_len = model_ctx_len
         self.fixed_ctx_len = fixed_ctx_len
+        self.fast_mc = fast_mc
         self.log_instances = 0  # Set to > 0 to log the first few instances as a sanity check
 
         self.samples: List[Dict[str, Any]] = []
@@ -1500,6 +1568,8 @@ class OEEvalTask(ICLMultiChoiceTaskDataset):
         for requests in self.dataset:
             current_doc_id_offset += max_doc_id
             max_doc_id = 0  # Max doc id seen in this dataset
+
+            new_samples = []
             for request in requests:
                 doc = request["doc"]
                 doc_id = request["doc_id"]
@@ -1571,7 +1641,7 @@ class OEEvalTask(ICLMultiChoiceTaskDataset):
                 dc_query = dc + continuation[:-1]
 
                 # form a sample
-                self.samples.append(
+                new_samples.append(
                     {
                         "doc_id": doc_id + current_doc_id_offset,
                         "cont_id": cont_id,
@@ -1592,6 +1662,46 @@ class OEEvalTask(ICLMultiChoiceTaskDataset):
                     }
                 )
 
+            # Fast MCQA:
+            # Only pass a single request, and group together all continuations as tokens
+            if self.fast_mc:
+                # Get unique doc IDs
+                unique_doc_ids = set(sample["doc_id"] for sample in new_samples)
+
+                # Create new samples list for fast MC
+                fast_mc_samples = []
+
+                # Process each unique document
+                for doc_id in unique_doc_ids:
+                    # Get all samples for this doc_id
+                    doc_samples = [s for s in new_samples if s["doc_id"] == doc_id]
+
+                    # Sort by continuation ID
+                    doc_samples.sort(key=lambda x: x["cont_id"])
+
+                    # Create new sample with distractor continuations
+                    base_sample = doc_samples[0].copy()
+                    choices = [s["continuation"] for s in doc_samples]
+
+                    # Assert all continuations are length 1
+                    for choice in choices:
+                        assert (
+                            len(choice) == 1
+                        ), f"Expected continuation length 1, got {len(choice)}"
+
+                    # Take first token of each continuation
+                    choices = [choice[0] for choice in choices]
+
+                    base_sample["choices"] = choices
+                    base_sample["fast_mc"] = True
+
+                    fast_mc_samples.append(base_sample)
+
+                # Add fast MC samples to main samples list
+                new_samples = fast_mc_samples
+
+            self.samples = new_samples
+
     def doc_to_text(self, doc) -> str:
         del doc
         raise NotImplementedError
@@ -1768,6 +1878,24 @@ LABEL_TO_TASK_MAP_ORIG = {
         OEEvalTask,
         {"dataset_path": "copycolors", "dataset_name": "xl_10way", "metric_type": "acc"},
     ),
+    "copycolors_10way_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "copycolors",
+            "dataset_name": "10way",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
+    "copycolors_xl_10way_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "copycolors",
+            "dataset_name": "xl_10way",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "csqa_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "csqa", "dataset_name": "mc_5shot", "metric_type": "acc"},
@@ -1792,6 +1920,10 @@ LABEL_TO_TASK_MAP_ORIG = {
         OEEvalTask,
         {"dataset_path": "hellaswag", "dataset_name": "rc_5shot", "metric_type": "len_norm"},
     ),
+    "hellaswag_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "hellaswag", "dataset_name": "rc_5shot", "metric_type": "bpb"},
+    ),
     "openbookqa_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "openbookqa", "dataset_name": "mc_5shot", "metric_type": "acc"},
@@ -2001,6 +2133,14 @@ LABEL_TO_TASK_MAP_LADDER = {
             "metric_type": "len_norm",
         },
     ),
+    "arc_challenge_val_bpb_5shot": (
+        OEEvalTask,
+        {
+            "dataset_path": "arc_challenge",
+            "dataset_name": "val_rc_5shot",
+            "metric_type": "bpb",
+        },
+    ),
     "arc_challenge_val_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "arc_challenge", "dataset_name": "val_mc_5shot", "metric_type": "acc"},
@@ -2013,114 +2153,299 @@ LABEL_TO_TASK_MAP_LADDER = {
             "metric_type": "len_norm",
         },
     ),
+    "arc_challenge_test_bpb_5shot": (
+        OEEvalTask,
+        {
+            "dataset_path": "arc_challenge",
+            "dataset_name": "test_rc_5shot",
+            "metric_type": "bpb",
+        },
+    ),
     "arc_challenge_test_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "arc_challenge", "dataset_name": "test_mc_5shot", "metric_type": "acc"},
     ),
+    "arc_challenge_test_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "arc_challenge",
+            "dataset_name": "test_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "arc_easy_val_rc_5shot": (
         OEEvalTask,
         {"dataset_path": "arc_easy", "dataset_name": "val_rc_5shot", "metric_type": "len_norm"},
     ),
+    "arc_easy_val_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "arc_easy", "dataset_name": "val_rc_5shot", "metric_type": "bpb"},
+    ),
     "arc_easy_val_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "arc_easy", "dataset_name": "val_mc_5shot", "metric_type": "acc"},
     ),
+    "arc_easy_val_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "arc_easy",
+            "dataset_name": "val_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "arc_easy_test_rc_5shot": (
         OEEvalTask,
         {"dataset_path": "arc_easy", "dataset_name": "test_rc_5shot", "metric_type": "len_norm"},
     ),
+    "arc_easy_test_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "arc_easy", "dataset_name": "test_rc_5shot", "metric_type": "bpb"},
+    ),
     "arc_easy_test_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "arc_easy", "dataset_name": "test_mc_5shot", "metric_type": "acc"},
     ),
+    "arc_easy_test_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "arc_easy",
+            "dataset_name": "test_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "boolq_val_rc_5shot": (
         OEEvalTask,
         {"dataset_path": "boolq", "dataset_name": "val_rc_5shot", "metric_type": "acc"},
     ),
+    "boolq_val_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "boolq", "dataset_name": "val_rc_5shot", "metric_type": "bpb"},
+    ),
     "boolq_val_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "boolq", "dataset_name": "val_mc_5shot", "metric_type": "acc"},
     ),
+    "boolq_val_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "boolq",
+            "dataset_name": "val_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "csqa_val_rc_5shot": (
         OEEvalTask,
         {"dataset_path": "csqa", "dataset_name": "val_rc_5shot", "metric_type": "len_norm"},
     ),
+    "csqa_val_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "csqa", "dataset_name": "val_rc_5shot", "metric_type": "bpb"},
+    ),
     "csqa_val_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "csqa", "dataset_name": "val_mc_5shot", "metric_type": "acc"},
     ),
+    "csqa_val_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "csqa",
+            "dataset_name": "val_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "hellaswag_val_rc_5shot": (
         OEEvalTask,
         {"dataset_path": "hellaswag", "dataset_name": "val_rc_5shot", "metric_type": "len_norm"},
     ),
+    "hellaswag_val_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "hellaswag", "dataset_name": "val_rc_5shot", "metric_type": "bpb"},
+    ),
     "hellaswag_val_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "hellaswag", "dataset_name": "val_mc_5shot", "metric_type": "acc"},
     ),
+    "hellaswag_val_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "hellaswag",
+            "dataset_name": "val_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "openbookqa_val_rc_5shot": (
         OEEvalTask,
         {"dataset_path": "openbookqa", "dataset_name": "val_rc_5shot", "metric_type": "len_norm"},
     ),
+    "openbookqa_val_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "openbookqa", "dataset_name": "val_rc_5shot", "metric_type": "bpb"},
+    ),
     "openbookqa_val_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "openbookqa", "dataset_name": "val_mc_5shot", "metric_type": "acc"},
     ),
+    "openbookqa_val_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "openbookqa",
+            "dataset_name": "val_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "openbookqa_test_rc_5shot": (
         OEEvalTask,
         {"dataset_path": "openbookqa", "dataset_name": "test_rc_5shot", "metric_type": "len_norm"},
     ),
+    "openbookqa_test_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "openbookqa", "dataset_name": "test_rc_5shot", "metric_type": "bpb"},
+    ),
     "openbookqa_test_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "openbookqa", "dataset_name": "test_mc_5shot", "metric_type": "acc"},
     ),
+    "openbookqa_test_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "openbookqa",
+            "dataset_name": "test_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "piqa_val_rc_5shot": (
         OEEvalTask,
         {"dataset_path": "piqa", "dataset_name": "val_rc_5shot", "metric_type": "len_norm"},
     ),
+    "piqa_val_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "piqa", "dataset_name": "val_rc_5shot", "metric_type": "bpb"},
+    ),
     "piqa_val_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "piqa", "dataset_name": "val_mc_5shot", "metric_type": "acc"},
     ),
+    "piqa_val_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "piqa",
+            "dataset_name": "val_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "socialiqa_val_rc_5shot": (
         OEEvalTask,
         {"dataset_path": "socialiqa", "dataset_name": "val_rc_5shot", "metric_type": "len_norm"},
     ),
+    "socialiqa_val_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "socialiqa", "dataset_name": "val_rc_5shot", "metric_type": "bpb"},
+    ),
     "socialiqa_val_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "socialiqa", "dataset_name": "val_mc_5shot", "metric_type": "acc"},
     ),
+    "socialiqa_val_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "socialiqa",
+            "dataset_name": "val_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "winogrande_val_rc_5shot": (
         OEEvalTask,
         {"dataset_path": "winogrande", "dataset_name": "val_rc_5shot", "metric_type": "len_norm"},
     ),
+    "winogrande_val_bpb_5shot": (
+        OEEvalTask,
+        {"dataset_path": "winogrande", "dataset_name": "val_rc_5shot", "metric_type": "bpb"},
+    ),
     "winogrande_val_mc_5shot": (
         OEEvalTask,
         {"dataset_path": "winogrande", "dataset_name": "val_mc_5shot", "metric_type": "acc"},
     ),
+    "winogrande_val_mc_5shot_fast": (
+        OEEvalTask,
+        {
+            "dataset_path": "winogrande",
+            "dataset_name": "val_mc_5shot",
+            "metric_type": "acc",
+            "fast_mc": True,
+        },
+    ),
     "mmlu_stem_val_rc_var": (MMLU, {"dataset_name": "stem", "prompt_variations": 1}),
     "mmlu_stem_val_rc_5shot": (MMLU, {"dataset_name": "stem", "prompt_variations": 2}),
+    "mmlu_stem_val_bpb_5shot": (
+        MMLU,
+        {"dataset_name": "stem", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
     "mmlu_stem_val_mc_5shot": (
         MMLU,
         {"dataset_name": "stem", "prompt_variations": 2, "mc_labels": True},
     ),
+    "mmlu_stem_val_mc_5shot_fast": (
+        MMLU,
+        {"dataset_name": "stem", "prompt_variations": 2, "mc_labels": True, "fast_mc": True},
+    ),
     "mmlu_stem_test_rc_var": (
         MMLU,
         {"dataset_name": "stem", "split": "test", "prompt_variations": 1},
     ),
+    "mmlu_stem_test_bpb_var": (
+        MMLU,
+        {"dataset_name": "stem", "split": "test", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
     "mmlu_stem_test_rc_5shot": (
         MMLU,
         {"dataset_name": "stem", "split": "test", "prompt_variations": 2},
     ),
+    "mmlu_stem_test_bpb_5shot": (
+        MMLU,
+        {"dataset_name": "stem", "split": "test", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
     "mmlu_stem_test_mc_5shot": (
         MMLU,
         {"dataset_name": "stem", "split": "test", "prompt_variations": 2, "mc_labels": True},
     ),
+    "mmlu_stem_test_mc_5shot_fast": (
+        MMLU,
+        {
+            "dataset_name": "stem",
+            "split": "test",
+            "prompt_variations": 2,
+            "mc_labels": True,
+            "fast_mc": True,
+        },
+    ),
     "mmlu_humanities_val_rc_var": (MMLU, {"dataset_name": "humanities", "prompt_variations": 1}),
     "mmlu_humanities_val_rc_5shot": (MMLU, {"dataset_name": "humanities", "prompt_variations": 2}),
+    "mmlu_humanities_val_bpb_var": (
+        MMLU,
+        {"dataset_name": "humanities", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
+    "mmlu_humanities_val_bpb_5shot": (
+        MMLU,
+        {"dataset_name": "humanities", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
     "mmlu_humanities_val_mc_5shot": (
         MMLU,
         {"dataset_name": "humanities", "prompt_variations": 2, "mc_labels": True},
     ),
+    "mmlu_humanities_val_mc_5shot_fast": (
+        MMLU,
+        {"dataset_name": "humanities", "prompt_variations": 2, "mc_labels": True, "fast_mc": True},
+    ),
     "mmlu_humanities_test_rc_var": (
         MMLU,
         {"dataset_name": "humanities", "split": "test", "prompt_variations": 1},
@@ -2129,10 +2454,38 @@ LABEL_TO_TASK_MAP_LADDER = {
         MMLU,
         {"dataset_name": "humanities", "split": "test", "prompt_variations": 2},
     ),
+    "mmlu_humanities_test_bpb_var": (
+        MMLU,
+        {
+            "dataset_name": "humanities",
+            "split": "test",
+            "prompt_variations": 2,
+            "metric_type": "bpb",
+        },
+    ),
+    "mmlu_humanities_test_bpb_5shot": (
+        MMLU,
+        {
+            "dataset_name": "humanities",
+            "split": "test",
+            "prompt_variations": 2,
+            "metric_type": "bpb",
+        },
+    ),
     "mmlu_humanities_test_mc_5shot": (
         MMLU,
         {"dataset_name": "humanities", "split": "test", "prompt_variations": 2, "mc_labels": True},
     ),
+    "mmlu_humanities_test_mc_5shot_fast": (
+        MMLU,
+        {
+            "dataset_name": "humanities",
+            "split": "test",
+            "prompt_variations": 2,
+            "mc_labels": True,
+            "fast_mc": True,
+        },
+    ),
     "mmlu_social_sciences_val_rc_var": (
         MMLU,
         {"dataset_name": "social_sciences", "prompt_variations": 1},
@@ -2141,10 +2494,27 @@ LABEL_TO_TASK_MAP_LADDER = {
         MMLU,
         {"dataset_name": "social_sciences", "prompt_variations": 2},
     ),
+    "mmlu_social_sciences_val_bpb_var": (
+        MMLU,
+        {"dataset_name": "social_sciences", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
+    "mmlu_social_sciences_val_bpb_5shot": (
+        MMLU,
+        {"dataset_name": "social_sciences", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
     "mmlu_social_sciences_val_mc_5shot": (
         MMLU,
         {"dataset_name": "social_sciences", "prompt_variations": 2, "mc_labels": True},
     ),
+    "mmlu_social_sciences_val_mc_5shot_fast": (
+        MMLU,
+        {
+            "dataset_name": "social_sciences",
+            "prompt_variations": 2,
+            "mc_labels": True,
+            "fast_mc": True,
+        },
+    ),
     "mmlu_social_sciences_test_rc_var": (
         MMLU,
         {"dataset_name": "social_sciences", "split": "test", "prompt_variations": 1},
@@ -2153,6 +2523,24 @@ LABEL_TO_TASK_MAP_LADDER = {
         MMLU,
         {"dataset_name": "social_sciences", "split": "test", "prompt_variations": 2},
     ),
+    "mmlu_social_sciences_test_bpb_var": (
+        MMLU,
+        {
+            "dataset_name": "social_sciences",
+            "split": "test",
+            "prompt_variations": 2,
+            "metric_type": "bpb",
+        },
+    ),
+    "mmlu_social_sciences_test_bpb_5shot": (
+        MMLU,
+        {
+            "dataset_name": "social_sciences",
+            "split": "test",
+            "prompt_variations": 2,
+            "metric_type": "bpb",
+        },
+    ),
     "mmlu_social_sciences_test_mc_5shot": (
         MMLU,
         {
@@ -2162,12 +2550,34 @@ LABEL_TO_TASK_MAP_LADDER = {
             "mc_labels": True,
         },
     ),
+    "mmlu_social_sciences_test_mc_5shot_fast": (
+        MMLU,
+        {
+            "dataset_name": "social_sciences",
+            "split": "test",
+            "prompt_variations": 2,
+            "mc_labels": True,
+            "fast_mc": True,
+        },
+    ),
     "mmlu_other_val_rc_var": (MMLU, {"dataset_name": "other", "prompt_variations": 1}),
     "mmlu_other_val_rc_5shot": (MMLU, {"dataset_name": "other", "prompt_variations": 2}),
+    "mmlu_other_val_bpb_var": (
+        MMLU,
+        {"dataset_name": "other", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
+    "mmlu_other_val_bpb_5shot": (
+        MMLU,
+        {"dataset_name": "other", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
     "mmlu_other_val_mc_5shot": (
         MMLU,
         {"dataset_name": "other", "prompt_variations": 2, "mc_labels": True},
     ),
+    "mmlu_other_val_mc_5shot_fast": (
+        MMLU,
+        {"dataset_name": "other", "prompt_variations": 2, "mc_labels": True, "fast_mc": True},
+    ),
     "mmlu_other_test_rc_var": (
         MMLU,
         {"dataset_name": "other", "split": "test", "prompt_variations": 1},
@@ -2176,10 +2586,28 @@ LABEL_TO_TASK_MAP_LADDER = {
         MMLU,
         {"dataset_name": "other", "split": "test", "prompt_variations": 2},
     ),
+    "mmlu_other_test_bpb_var": (
+        MMLU,
+        {"dataset_name": "other", "split": "test", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
+    "mmlu_other_test_bpb_5shot": (
+        MMLU,
+        {"dataset_name": "other", "split": "test", "prompt_variations": 2, "metric_type": "bpb"},
+    ),
     "mmlu_other_test_mc_5shot": (
         MMLU,
         {"dataset_name": "other", "split": "test", "prompt_variations": 2, "mc_labels": True},
     ),
+    "mmlu_other_test_mc_5shot_fast": (
+        MMLU,
+        {
+            "dataset_name": "other",
+            "split": "test",
+            "prompt_variations": 2,
+            "mc_labels": True,
+            "fast_mc": True,
+        },
+    ),
 }
 
 # Expanded tasks for BPB on some generative tasks

olmo_eval/version.py

@@ -1,6 +1,6 @@
 _MAJOR = "0"
-_MINOR = "7"
-_PATCH = "2"
+_MINOR = "8"
+_PATCH = "0"
 _SUFFIX = ""
 
 VERSION_SHORT = "{0}.{1}".format(_MAJOR, _MINOR)