crfm-helm 0.5.3__py3-none-any.whl → 0.5.5__py3-none-any.whl

helm/benchmark/reeval_runner.py

@@ -0,0 +1,355 @@
+import dacite
+import json
+import os
+import typing
+from collections import Counter
+from typing import Any, Dict, List, Optional
+import torch
+
+from tqdm import tqdm
+from dataclasses import replace
+from datasets import load_dataset
+
+from helm.benchmark.adaptation.request_state import RequestState
+from helm.common.general import ensure_directory_exists, write, asdict_without_nones
+from helm.common.hierarchical_logger import hlog, htrack_block
+from helm.common.cache import cache_stats
+from helm.benchmark.scenarios.scenario import (
+    Scenario,
+    create_scenario,
+    Instance,
+    get_scenario_cache_path,
+    with_instance_ids,
+)
+from helm.benchmark.adaptation.adapters.adapter import Adapter
+from helm.benchmark.adaptation.adapters.adapter_factory import AdapterFactory
+from helm.benchmark.adaptation.scenario_state import ScenarioState
+from helm.benchmark.run_spec import RunSpec
+from helm.benchmark.data_preprocessor import DataPreprocessor
+from helm.benchmark.executor import ExecutionSpec
+from helm.benchmark.metrics.metric_name import MetricName
+from helm.benchmark.metrics.dry_run_metrics import DryRunMetric
+from helm.benchmark.metrics.metric import MetricInterface, MetricResult, PerInstanceStats, create_metric, Stat
+from helm.benchmark.runner import (
+    Runner,
+    remove_stats_nans,
+    remove_per_instance_stats_nans,
+)
+
+scenario_to_metric_name = {
+    "air_bench_2024": "air_score",
+    "babi_qa": "quasi_exact_match",
+    "bbq": "quasi_exact_match",
+    "blimp": "exact_match",
+    "boolq": "quasi_exact_match",
+    "civil_comments": "quasi_exact_match",
+    "dyck_language": "exact_match_indicator",
+    "entity_data_imputation": "quasi_exact_match",
+    "entity_matching": "quasi_exact_match",
+    "imdb": "quasi_exact_match",
+    "legal_support": "quasi_exact_match",
+    "raft": "quasi_exact_match",
+    "synthetic_reasoning": "quasi_exact_match",
+    "truthful_qa": "exact_match",
+    "wikifact": "quasi_exact_match",
+    "mmlu": "exact_match",
+    "commonsense": "exact_match",
+    "gsm": "final_number_exact_match",
+    # "gsm": ["exact_match_indicator", "final_number_exact_match"],
+    "legalbench": "quasi_exact_match",
+    "math": "math_equiv_chain_of_thought",
+    "med_qa": "quasi_exact_match",
+    "thai_exam": "exact_match",
+}
+
+
+class REEvalRunner(Runner):
+    """
+    This runner implements the basic (non-amortized) method described in the paper
+    `Reliable and Efficient Amortized Model-Based Evaluation`. This approach, which is
+    also known as Computerized Adaptive Testing (CAT) within the framework of Item Response
+    Theory (IRT), leverages adaptive testing to evaluate model performance.
+
+    The difficulties of the questions are provided in a HuggingFace repository. In addition,
+    the authors of the paper will supply a Python package for calculating these difficulties.
+    At each iteration, the runner estimates the model's ability based on all previously
+    administered questions and their corresponding responses. It then selects the next question
+    whose difficulty is closest to the estimated ability, thereby reliably and efficiently
+    eliciting the model's ability.
+    """
+
+    def __init__(
+        self,
+        execution_spec: ExecutionSpec,
+        output_path: str,
+        suite: str,
+        skip_instances: bool,
+        cache_instances: bool,
+        cache_instances_only: bool,
+        skip_completed_runs: bool,
+        exit_on_error: bool,
+    ):
+        super().__init__(
+            execution_spec=execution_spec,
+            output_path=output_path,
+            suite=suite,
+            skip_instances=skip_instances,
+            cache_instances=cache_instances,
+            cache_instances_only=cache_instances_only,
+            skip_completed_runs=skip_completed_runs,
+            exit_on_error=exit_on_error,
+        )
+
+    def _estimate_model_ability(
+        self,
+        old_ability: float,
+        response_correctness: List[float],
+        instance_difficulties: List[float],
+    ) -> float:
+        def closure():
+            optim.zero_grad()
+            probs = torch.sigmoid(ability + difficulties)
+            loss = -torch.distributions.Bernoulli(probs=probs).log_prob(responses).mean()
+            loss.backward()
+            return loss
+
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        responses = torch.tensor(response_correctness, device=device)
+        difficulties = torch.tensor(instance_difficulties, device=device)
+        ability = torch.tensor([old_ability], requires_grad=True, device=device)
+        optim = torch.optim.LBFGS([ability], lr=0.1, max_iter=20, history_size=10, line_search_fn="strong_wolfe")
+
+        for iteration in range(100):
+            loss = optim.step(closure)
+
+            if iteration > 0:
+                prev_ability = ability.clone()
+                prev_loss = loss
+                d_loss = prev_loss - loss
+                d_theta = torch.norm(prev_ability - ability, p=2)
+                grad_norm = torch.norm(optim.param_groups[0]["params"][0].grad, p=2)
+                if d_loss < 1e-5 and d_theta < 1e-5 and grad_norm < 1e-5:
+                    break
+
+        return ability.item()
+
+    def run_one(self, run_spec: RunSpec):
+        run_path: str = self._get_run_path(run_spec)
+        if self.skip_completed_runs and self._is_run_completed(run_path):
+            hlog(f"Skipping run {run_spec.name} because run is completed and all output files exist.")
+            return
+        ensure_directory_exists(run_path)
+
+        # Load the scenario
+        scenario: Scenario = create_scenario(run_spec.scenario_spec)
+
+        # This 'output_path' will be used when the model's input instances are saved.
+        args_str = ",".join([f"{k}={v}" for k, v in sorted(run_spec.scenario_spec.args.items())])
+        scenario_name_with_args = f"{scenario.name}:{args_str}" if args_str else f"{scenario.name}"
+        input_instances_output_path = os.path.join(self.instances_path, scenario_name_with_args)
+        input_instances_file_path = os.path.join(input_instances_output_path, "input_instances.json")
+
+        instances: List[Instance]
+        if self.skip_instances:
+            instances = []
+        else:
+            if self.cache_instances and os.path.exists(input_instances_file_path):
+                with open(input_instances_file_path) as f:
+                    json_instances: List[Dict[str, Any]] = json.load(f)
+                instances = [dacite.from_dict(Instance, instance) for instance in json_instances]
+            else:
+                # Create the instances of the scenario
+                scenario_output_path = get_scenario_cache_path(self.output_path, scenario.name)
+                with htrack_block("scenario.get_instances"):
+                    instances = scenario.get_instances(scenario_output_path)
+        if self.cache_instances and not os.path.exists(input_instances_file_path):
+            # Save instances to file
+            ensure_directory_exists(input_instances_output_path)
+            write(
+                os.path.join(input_instances_file_path),
+                json.dumps([asdict_without_nones(instance) for instance in instances], indent=2),
+            )
+        if self.cache_instances_only:
+            return  # Exit after saving the instances.
+
+        # Give each instance a unique ID
+        if any([instance.id is None for instance in instances]):
+            instances = with_instance_ids(instances)
+
+        # Data preprocessing
+        instances = DataPreprocessor(run_spec.data_augmenter_spec).preprocess(
+            instances, self.executor.execution_spec.parallelism
+        )
+
+        # Adapt (convert to requests)
+        adapter: Adapter = AdapterFactory.get_adapter(run_spec.adapter_spec, self.tokenizer_service)
+        unasked_request_states_without_difficulty: List[RequestState] = adapter.adapt(
+            instances, self.executor.execution_spec.parallelism
+        )
+
+        # load difficulty
+        split_name = "dyck_language_np_3" if scenario.name == "dyck_language" else scenario.name
+        try:
+            difficulty_dataset = load_dataset("stair-lab/reeval-difficulty", split=split_name)
+            prompt_to_difficulty: dict[str, float] = {row["request.prompt"]: row["z"] for row in difficulty_dataset}
+        except ValueError:
+            hlog(f"WARNING: no available difficulty for {split_name}, skipping")
+            return
+
+        unasked_request_states: List[RequestState] = []
+        for request_state in unasked_request_states_without_difficulty:
+            prompt = request_state.request.prompt
+            if prompt in prompt_to_difficulty:
+                difficulty = prompt_to_difficulty[prompt]
+                current_extra_data = request_state.instance.extra_data or {}
+                if "difficulty" in current_extra_data:
+                    raise Exception("Extra_data already contains a 'difficulty' key.")
+                new_extra_data = current_extra_data.copy()
+                new_extra_data["difficulty"] = difficulty
+                new_instance = replace(request_state.instance, extra_data=new_extra_data)
+                new_request_state = replace(request_state, instance=new_instance)
+                unasked_request_states.append(new_request_state)
+        assert unasked_request_states
+
+        # Execute the requests in an reeval manner
+        assert run_spec.adapter_spec.reeval_parameters is not None
+        model_ability: float = run_spec.adapter_spec.reeval_parameters.model_ability or 0.0
+        scenario_metric_name: str = scenario_to_metric_name[scenario.name]
+
+        asked_request_states: List[RequestState] = []
+        reeval_trajectory: Dict[str, List[float]] = {
+            "model_ability": [],
+            "response_correctness": [],
+            "instance_difficulties": [],
+        }
+
+        assert run_spec.adapter_spec.max_eval_instances is not None
+        for _ in tqdm(range(run_spec.adapter_spec.max_eval_instances), desc="REEval Execution"):
+            if not unasked_request_states:
+                break
+
+            selected_item: Optional[RequestState] = None
+            min_diff = float("inf")
+            for item in unasked_request_states:
+                assert item.instance.extra_data is not None
+                diff = abs(item.instance.extra_data["difficulty"] + model_ability)
+                if diff < min_diff:
+                    min_diff = diff
+                    selected_item = item
+            assert selected_item is not None
+            unasked_request_states.remove(selected_item)
+
+            # Execute the request
+            single_scenario_state: ScenarioState = ScenarioState(
+                adapter_spec=run_spec.adapter_spec,
+                request_states=[selected_item],
+                annotator_specs=run_spec.annotators,
+            )
+
+            # Execute (fill up results)
+            single_scenario_state = self.executor.execute(single_scenario_state)
+
+            # Annotate (post-process the results)
+            single_scenario_state = self.annotator_executor.execute(single_scenario_state)
+
+            # Apply the metrics
+            # When performing a dry run, only estimate the number of tokens instead
+            # of calculating the metrics.
+            metrics: List[MetricInterface] = (
+                [DryRunMetric()]
+                if self.dry_run
+                else [create_metric(metric_spec) for metric_spec in run_spec.metric_specs]
+            )
+
+            temp_per_instance_stats: List[PerInstanceStats] = []
+            with htrack_block(f"{len(metrics)} metrics"):
+                for metric in metrics:
+                    with htrack_block(metric):
+                        temp_metric_result: MetricResult = metric.evaluate(
+                            single_scenario_state,
+                            self.metric_service,
+                            self.eval_cache_path,
+                            self.executor.execution_spec.parallelism,
+                        )
+                        temp_per_instance_stats.extend(temp_metric_result.per_instance_stats)
+
+            # Update the reeval request states
+            asked_request_states.extend(single_scenario_state.request_states)
+
+            # Update the reeval trajectory
+            reeval_trajectory["model_ability"].append(model_ability)
+            scenario_metric_value = [
+                s for s in temp_per_instance_stats[0].stats if s.name.name == scenario_metric_name
+            ][0].mean
+
+            assert scenario_metric_value is not None
+            reeval_trajectory["response_correctness"].append(scenario_metric_value)
+            assert selected_item.instance.extra_data is not None
+            reeval_trajectory["instance_difficulties"].append(selected_item.instance.extra_data["difficulty"])
+
+            # Estimate the model ability
+            model_ability = self._estimate_model_ability(
+                old_ability=model_ability,
+                response_correctness=reeval_trajectory["response_correctness"],
+                instance_difficulties=reeval_trajectory["instance_difficulties"],
+            )
+
+        # Create the scenario state
+        scenario_state: ScenarioState = ScenarioState(
+            adapter_spec=run_spec.adapter_spec,
+            request_states=asked_request_states,
+            annotator_specs=run_spec.annotators,
+        )
+
+        stats: List[Stat] = []
+        per_instance_stats: List[PerInstanceStats] = []
+        with htrack_block(f"{len(metrics)} metrics"):
+            for metric in metrics:
+                with htrack_block(metric):
+                    metric_result: MetricResult = metric.evaluate(
+                        scenario_state,
+                        self.metric_service,
+                        self.eval_cache_path,
+                        self.executor.execution_spec.parallelism,
+                    )
+                    stats.extend(metric_result.aggregated_stats)
+                    per_instance_stats.extend(metric_result.per_instance_stats)
+
+        # Check that there aren't duplicate `Stat`s
+        # Note: doesn't catch near misses.
+        metric_counts: typing.Counter[MetricName] = Counter([stat.name for stat in stats])
+        for metric_name, count in metric_counts.items():
+            if count > 1:
+                hlog(f"WARNING: duplicate metric name {metric_name}")
+
+        # Print out the number of stats
+        hlog(f"Generated {len(stats)} stats.")
+
+        if self.skip_instances:
+            hlog("skip_instances was True. Skipping writing results out.")
+            return
+
+        # Output benchmarking information and results to files
+        write(os.path.join(run_path, "run_spec.json"), json.dumps(asdict_without_nones(run_spec), indent=2))
+
+        # Write out scenario
+        write(os.path.join(run_path, "scenario.json"), json.dumps(asdict_without_nones(scenario), indent=2))
+
+        # Write scenario state
+        write(os.path.join(run_path, "scenario_state.json"), json.dumps(asdict_without_nones(scenario_state), indent=2))
+
+        write(
+            os.path.join(run_path, "stats.json"),
+            json.dumps([asdict_without_nones(stat) for stat in remove_stats_nans(stats)], indent=2),
+        )
+        write(
+            os.path.join(run_path, "per_instance_stats.json"),
+            json.dumps(list(map(asdict_without_nones, remove_per_instance_stats_nans(per_instance_stats))), indent=2),
+        )
+
+        write(
+            os.path.join(run_path, "reeval_trajectory.json"),
+            json.dumps(reeval_trajectory, indent=2),
+        )
+
+        cache_stats.print_status()

helm/benchmark/run.py

@@ -266,13 +266,6 @@ def main():
         default=None,
         help="Full class name of the Runner class to use. If unset, uses the default Runner.",
     )
-    parser.add_argument(
-        "--openvino",
-        action="store_true",
-        default=False,
-        help="Experimental: Apply openvino optimization to Hugging Face AutoModelForCausalLM models "
-        "specified with the --enable-huggingface-models and --enable-local-huggingface-models flags.",
-    )
     add_run_args(parser)
     args = parser.parse_args()
     validate_args(args)
@@ -284,19 +277,13 @@ def main():
         from helm.benchmark.huggingface_registration import register_huggingface_hub_model_from_flag_value
 
         for huggingface_model_name in args.enable_huggingface_models:
-            if args.openvino:
-                register_huggingface_hub_model_from_flag_value(huggingface_model_name, args.openvino)
-            else:
-                register_huggingface_hub_model_from_flag_value(huggingface_model_name)
+            register_huggingface_hub_model_from_flag_value(huggingface_model_name)
 
     if args.enable_local_huggingface_models:
         from helm.benchmark.huggingface_registration import register_huggingface_local_model_from_flag_value
 
         for huggingface_model_path in args.enable_local_huggingface_models:
-            if args.openvino:
-                register_huggingface_local_model_from_flag_value(huggingface_model_path, args.openvino)
-            else:
-                register_huggingface_local_model_from_flag_value(huggingface_model_path)
+            register_huggingface_local_model_from_flag_value(huggingface_model_path)
 
     run_entries: List[RunEntry] = []
     if args.conf_paths:
@@ -323,12 +310,16 @@ def main():
             if model_to_run not in all_models:
                 raise Exception(f"Unknown model '{model_to_run}' passed to --models-to-run")
     else:
-        model_expander_pattern = re.compile(
+        model_expander_wildcard_pattern = re.compile(
             r"\bmodel=(?:all|text_code|text|code|instruction_following|full_functionality_text|limited_functionality_text)\b"  # noqa: E501
         )
-        if any(model_expander_pattern.search(run_entry.description) for run_entry in run_entries):
+        if any(model_expander_wildcard_pattern.search(run_entry.description) for run_entry in run_entries):
             raise Exception("--models-to-run must be set if the `models=` run expander expands to multiple models")
 
+        model_expander_pattern = re.compile(r"\bmodel=\b")
+        if not any(model_expander_pattern.search(run_entry.description) for run_entry in run_entries):
+            raise Exception("--models-to-run must be set if the `models=` run expander is omitted")
+
     run_specs = run_entries_to_run_specs(
         run_entries=run_entries,
         max_eval_instances=args.max_eval_instances,

helm/benchmark/run_expander.py

@@ -3,6 +3,7 @@ from abc import ABC, abstractmethod
 from dataclasses import replace
 from typing import Any, List, Dict, Optional, Tuple, Type
 
+from helm.benchmark.metrics.metric import MetricSpec
 from helm.benchmark.model_metadata_registry import (
     get_all_instruction_following_models,
     get_all_code_models,
@@ -11,19 +12,21 @@ from helm.benchmark.model_metadata_registry import (
     get_model_metadata,
     get_model_names_with_tag,
     DEPRECATED_MODEL_TAG,
+    UNSUPPORTED_MODEL_TAG,
     FULL_FUNCTIONALITY_TEXT_MODEL_TAG,
     LIMITED_FUNCTIONALITY_TEXT_MODEL_TAG,
     ABLATION_MODEL_TAG,
     TEXT_TO_IMAGE_MODEL_TAG,
     VISION_LANGUAGE_MODEL_TAG,
+    AUDIO_LANGUAGE_MODEL_TAG,
     INSTRUCTION_FOLLOWING_MODEL_TAG,
 )
 from helm.benchmark.adaptation.adapters.adapter_factory import ADAPT_GENERATION
 from helm.benchmark.model_deployment_registry import get_model_names_with_tokenizer
-from .run_spec import RunSpec
+from helm.benchmark.run_spec import RunSpec
 from helm.benchmark.adaptation.adapter_spec import ADAPT_MULTIPLE_CHOICE_JOINT, AdapterSpec, Substitution
-from .augmentations.perturbation import PerturbationSpec
-from .augmentations.data_augmenter import DataAugmenterSpec
+from helm.benchmark.augmentations.perturbation import PerturbationSpec
+from helm.benchmark.augmentations.data_augmenter import DataAugmenterSpec
 from helm.benchmark.scenarios.scenario import TEST_SPLIT, VALID_SPLIT
 
 
@@ -347,6 +350,29 @@ class AnthropicClaude3RunExpander(RunExpander):
         return [run_spec]
 
 
+class NovaRunExpander(RunExpander):
+    """
+    Custom prompt for Amazon Nova models.
+    These models need more explicit instructions about following the format.
+    """
+
+    name = "amazon-nova"
+
+    PROMPT = "Do not provide any additional explanation. Follow the format shown in the provided examples strictly."
+
+    def __init__(self):
+        pass
+
+    def expand(self, run_spec: RunSpec) -> List[RunSpec]:
+        return [
+            replace(
+                run_spec,
+                name=run_spec.name,
+                adapter_spec=replace(run_spec.adapter_spec, global_prefix=NovaRunExpander.PROMPT + "\n\n"),
+            ),
+        ]
+
+
 class FollowFormatInstructionsRunExpander(RunExpander):
     """Adds more explicit instructions about following the format to prompts.
 
@@ -588,6 +614,7 @@ class ModelRunExpander(ReplaceValueRunExpander):
             "opinions_qa_ai21": ["ai21/j1-grande", "ai21/j1-jumbo", "ai21/j1-grande-v2-beta"],
             "text_to_image": get_model_names_with_tag(TEXT_TO_IMAGE_MODEL_TAG),
             "vlm": get_model_names_with_tag(VISION_LANGUAGE_MODEL_TAG),
+            "audiolm": get_model_names_with_tag(AUDIO_LANGUAGE_MODEL_TAG),
         }
 
         # For each of the keys above (e.g., "text"), create a corresponding ablation (e.g., "ablation_text")
@@ -604,8 +631,10 @@ class ModelRunExpander(ReplaceValueRunExpander):
 
         # For each of the keys above, filter out deprecated models.
         deprecated_models = set(get_model_names_with_tag(DEPRECATED_MODEL_TAG))
+        unsupported_models = set(get_model_names_with_tag(UNSUPPORTED_MODEL_TAG))
+        excluded_models = deprecated_models | unsupported_models
         for family_name in values_dict.keys():
-            values_dict[family_name] = [model for model in values_dict[family_name] if model not in deprecated_models]
+            values_dict[family_name] = [model for model in values_dict[family_name] if model not in excluded_models]
 
         return values_dict
 
@@ -1271,6 +1300,30 @@ class IncreaseMaxTokensRunExpander(RunExpander):
         ]
 
 
+class TemperatureRunExpander(RunExpander):
+    """
+    Run expander for setting the temperature.
+    """
+
+    name = "temperature"
+
+    def __init__(self, value: float):
+        """
+        Args:
+            value (float): The amount to set temperature to
+        """
+        self.value = value
+
+    def expand(self, run_spec: RunSpec) -> List[RunSpec]:
+        adapter_spec = replace(run_spec.adapter_spec, temperature=self.value)
+        return [
+            replace(
+                run_spec,
+                adapter_spec=adapter_spec,
+            ),
+        ]
+
+
 class IncreaseTemperatureRunExpander(RunExpander):
     """
     Run expander for increasing the temperature.
@@ -1400,14 +1453,20 @@ class OutputFormatInstructions(RunExpander):
     name = "output_format_instructions"
 
     _SUFFIX_SUFFIX = "_suffix"
+    _NO_PREFIX_SUFFIX = "_no_prefix"
 
     def __init__(self, scenario: str):
+        self.suffix = False
         if scenario.endswith(OutputFormatInstructions._SUFFIX_SUFFIX):
-            self.scenario = scenario[: -len(OutputFormatInstructions._SUFFIX_SUFFIX)]
+            scenario = scenario.removesuffix(OutputFormatInstructions._SUFFIX_SUFFIX)
             self.suffix = True
-        else:
-            self.scenario = scenario
-            self.suffix = False
+
+        self.no_prefix = False
+        if scenario.endswith(OutputFormatInstructions._NO_PREFIX_SUFFIX):
+            scenario = scenario.removesuffix(OutputFormatInstructions._NO_PREFIX_SUFFIX)
+            self.no_prefix = True
+
+        self.scenario = scenario
 
     def expand(self, run_spec: RunSpec) -> List[RunSpec]:
         if run_spec.adapter_spec.method == ADAPT_MULTIPLE_CHOICE_JOINT:
@@ -1417,6 +1476,8 @@ class OutputFormatInstructions(RunExpander):
                 instructions = "Answer with only a single letter."
             elif self.scenario == "mcqa":
                 instructions = "Answer with only a single letter."
+            elif self.scenario == "mcqa_only_last_question":
+                instructions = "Answer only the last question with only a single letter."
             else:
                 instructions = "Answer with only a single letter."
         elif run_spec.adapter_spec.method == ADAPT_GENERATION:
@@ -1428,6 +1489,8 @@ class OutputFormatInstructions(RunExpander):
                 )
             elif self.scenario == "natural_qa":
                 instructions = "Answer with a short answer or a boolean 'yes' or 'no' answer."
+            elif self.scenario == "natural_qa_short_answer":
+                instructions = "Answer with a short answer."
             elif self.scenario == "legalbench":
                 if output_noun != "Answer":
                     instructions = f"Answer with the {output_noun.lower()}."
@@ -1459,6 +1522,11 @@ class OutputFormatInstructions(RunExpander):
             else:
                 raise ValueError(f"Unknown scenario {self.scenario}")
 
+        if self.no_prefix:
+            if instructions:
+                instructions += " "
+            instructions += f"Do not include '{run_spec.adapter_spec.output_prefix.strip()}' in your answer."
+
         if self.suffix:
             return [
                 replace(
@@ -1482,6 +1550,31 @@ class OutputFormatInstructions(RunExpander):
         ]
 
 
+class ProcessOutputRunExpander(RunExpander):
+    name = "process_output"
+
+    def __init__(self, processor: str):
+        self.processor = processor
+
+    def expand(self, run_spec: RunSpec) -> List[RunSpec]:
+        output_processing_metric_spec = MetricSpec(
+            class_name="helm.benchmark.metrics.output_processing_metric.OutputProcessingMetric",
+            args={
+                "processor": self.processor,
+                "metric_specs": [
+                    {"class_name": metric_spec.class_name, "args": metric_spec.args}
+                    for metric_spec in run_spec.metric_specs
+                ],
+            },
+        )
+        return [
+            replace(
+                run_spec,
+                metric_specs=[output_processing_metric_spec],
+            ),
+        ]
+
+
 RUN_EXPANDER_SUBCLASSES: List[Type[RunExpander]] = [
     InstructionsRunExpander,
     PromptRunExpander,
@@ -1505,6 +1598,10 @@ RUN_EXPANDER_SUBCLASSES: List[Type[RunExpander]] = [
     ChatMLRunExpander,
     EvalSplitRunExpander,
     OutputFormatInstructions,
+    TemperatureRunExpander,
+    IncreaseTemperatureRunExpander,
+    IncreaseMaxTokensRunExpander,
+    ProcessOutputRunExpander,
 ]
 
 

helm/benchmark/run_spec_factory.py

@@ -37,6 +37,7 @@ from helm.benchmark.run_expander import (
     IncreaseTemperatureRunExpander,
     IncreaseMaxTokensRunExpander,
     LlavaRunExpander,
+    ModelRunExpander,
     OpenFlamingoRunExpander,
     StopRunExpander,
 )
@@ -61,6 +62,10 @@ def construct_run_specs(spec: ObjectSpec) -> List[RunSpec]:
     expanders = [RUN_EXPANDERS[key](value) for key, value in args.items() if key in RUN_EXPANDERS]  # type: ignore
     args = dict((key, value) for key, value in args.items() if key not in RUN_EXPANDERS)
 
+    # If no model run expander was specified, add the model=all run expander
+    if not any([expander for expander in expanders if isinstance(expander, ModelRunExpander)]):
+        expanders.append(ModelRunExpander("all"))
+
     run_specs: List[RunSpec] = [run_spec_function(**args)]
 
     # Apply expanders
@@ -138,6 +143,13 @@ def construct_run_specs(spec: ObjectSpec) -> List[RunSpec]:
         ):
             run_spec = singleton(IncreaseMaxTokensRunExpander(value=1).expand(run_spec))
 
+        if model.name == "openai/o1-2024-12-17":
+            # From https://platform.openai.com/docs/guides/reasoning,
+            # "OpenAI recommends reserving at least 25,000 tokens for reasoning and outputs when you start
+            # experimenting with these models. As you become familiar with the number of reasoning tokens your
+            # prompts require, you can adjust this buffer accordingly."
+            run_spec = singleton(IncreaseMaxTokensRunExpander(value=25_000).expand(run_spec))
+
         # IDEFICS special handling
         if IDEFICS_MODEL_TAG in model.tags:
             if IDEFICS_INSTRUCT_MODEL_TAG in model.tags: