crfm-helm 0.4.0__py3-none-any.whl → 0.5.10__py3-none-any.whl

helm/benchmark/run_specs/classic_run_specs.py

@@ -0,0 +1,1393 @@
+"""Run spec functions for the HELM Classic leaderboard.
+
+Website: https://crfm.stanford.edu/helm/classic/
+
+If a run spec function is included in both the HELM Classic leaderboard and the
+HELM Lite leaderboard, it will be included in the lite_run_specs module instead of this module.
+This module also contains some scenarios that are currently not used on any HELM leaderboard."""
+
+from typing import Any, Dict, List, Optional, Set
+
+from helm.benchmark.adaptation.adapter_spec import (
+    ADAPT_GENERATION,
+    ADAPT_MULTIPLE_CHOICE_JOINT,
+    ADAPT_MULTIPLE_CHOICE_SEPARATE_ORIGINAL,
+    ADAPT_RANKING_BINARY,
+    AdapterSpec,
+)
+from helm.benchmark.adaptation.adapters.binary_ranking_adapter import BinaryRankingAdapter
+from helm.benchmark.adaptation.common_adapter_specs import (
+    get_completion_adapter_spec,
+    get_generation_adapter_spec,
+    get_language_modeling_adapter_spec,
+    get_multiple_choice_adapter_spec,
+    get_ranking_binary_adapter_spec,
+    get_summarization_adapter_spec,
+)
+from helm.benchmark.annotation.annotator import AnnotatorSpec
+from helm.benchmark.metrics.common_metric_specs import (
+    get_basic_metric_specs,
+    get_bias_metric_specs,
+    get_classification_metric_specs,
+    get_copyright_metric_specs,
+    get_disinformation_metric_specs,
+    get_exact_match_metric_specs,
+    get_f1_metric_specs,
+    get_generative_harms_metric_specs,
+    get_language_modeling_metric_specs,
+    get_open_ended_generation_metric_specs,
+    get_summarization_metric_specs,
+    get_basic_generation_metric_specs,
+    get_basic_reference_metric_specs,
+    get_generic_metric_specs,
+)
+from helm.benchmark.metrics.metric import MetricSpec
+from helm.benchmark.run_spec import RunSpec, run_spec_function
+from helm.benchmark.runner import get_benchmark_output_path
+from helm.benchmark.scenarios.scenario import ScenarioSpec, get_scenario_cache_path
+from helm.common.hierarchical_logger import hlog, htrack
+
+
+@run_spec_function("bbq")
+def get_bbq_spec(subject: str, method: str = ADAPT_MULTIPLE_CHOICE_JOINT) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.bbq_scenario.BBQScenario", args={"subject": subject}
+    )
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=method,
+        instructions="The following are multiple choice questions (with answers).",
+        input_noun="Passage",
+        output_noun="Answer",
+    )
+    metric_specs = [
+        MetricSpec(class_name="helm.benchmark.metrics.bbq_metrics.BBQMetric", args={})
+    ] + get_exact_match_metric_specs()
+
+    return RunSpec(
+        name=f"bbq:subject={subject},method={method}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=metric_specs,
+        groups=["bbq"],
+    )
+
+
+@run_spec_function("msmarco")
+def get_msmarco_spec(track: str, valid_topk: Optional[int] = None) -> RunSpec:
+    from helm.benchmark.scenarios.msmarco_scenario import MSMARCOScenario
+
+    valid_topk = None if valid_topk is None else int(valid_topk)
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.msmarco_scenario.MSMARCOScenario",
+        args={"track": track, "valid_topk": valid_topk},
+    )
+
+    adapter_spec: AdapterSpec = get_ranking_binary_adapter_spec(max_train_instances=4, stop_sequences=["\n"])
+
+    # Names of the measures we want to compute.
+    measure_names = MSMARCOScenario.MEASURE_NAMES[track]
+    multiple_relevance_values = set(MSMARCOScenario.GOLD_RELATIONS[track]) != {1}
+
+    metric_specs = (
+        [
+            MetricSpec(
+                class_name="helm.benchmark.metrics.ranking_metrics.RankingMetric",
+                args={
+                    "method": ADAPT_RANKING_BINARY,
+                    "measure_names": measure_names,
+                    "correct_output": BinaryRankingAdapter.RANKING_CORRECT_LABEL,
+                    "wrong_output": BinaryRankingAdapter.RANKING_WRONG_LABEL,
+                    "rank": valid_topk,
+                    "multiple_relevance_values": multiple_relevance_values,
+                },
+            ),
+        ]
+        + get_basic_reference_metric_specs()
+        + get_generic_metric_specs()
+    )
+
+    return RunSpec(
+        name=f"msmarco:track={track},valid_topk={valid_topk}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=metric_specs,
+        groups=[f"msmarco_{track}"],
+    )
+
+
+@run_spec_function("bold")
+def get_bold_spec(subject: str) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.bold_scenario.BOLDScenario", args={"subject": subject}
+    )
+
+    adapter_spec = get_completion_adapter_spec(
+        temperature=0.9,  # Set to approximate nucleus sampling conditions.
+        max_tokens=20,  # See Table 8 of RealToxicityPrompts: https://arxiv.org/pdf/2009.11462.pdf
+    )
+
+    return RunSpec(
+        name=f"bold:subject={subject}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_generative_harms_metric_specs(include_basic_metrics=True),
+        groups=["bold"],
+    )
+
+
+@run_spec_function("civil_comments")
+def get_civil_comments_spec(demographic: str) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.civil_comments_scenario.CivilCommentsScenario",
+        args={"demographic": demographic},
+    )
+
+    adapter_spec = get_generation_adapter_spec(input_noun="Passage", output_noun="Answer")
+
+    return RunSpec(
+        name=f"civil_comments:demographic={demographic}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs() + get_bias_metric_specs() + get_classification_metric_specs(),
+        groups=["civil_comments"],
+    )
+
+
+@run_spec_function("custom_mcqa")
+def get_custom_mcqa_spec(
+    path: str,
+    num_train_instances: int = 0,
+    method: str = ADAPT_MULTIPLE_CHOICE_JOINT,
+) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.custom_mcqa_scenario.CustomMCQAScenario",
+        args={
+            "path": path,
+            "num_train_instances": num_train_instances,
+        },
+    )
+
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=method,
+        instructions="The following are multiple choice questions (with answers).",
+        input_noun="Question",
+        output_noun="Answer",
+        max_train_instances=num_train_instances,
+    )
+
+    return RunSpec(
+        name=f"custom_mcqa,path={path},method={method}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs(),
+        groups=["custom"],
+    )
+
+
+@run_spec_function("interactive_qa_mmlu")
+def get_interactive_qa_mmlu_spec(subject: str) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.interactive_qa_mmlu_scenario.InteractiveQAMMLUScenario",
+        args={"subject": subject},
+    )
+
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=ADAPT_MULTIPLE_CHOICE_JOINT,
+        instructions=f"The following are multiple choice questions (with answers) about {subject.replace('_', ' ')}.",
+        input_noun="Question",
+        output_noun="Answer",
+    )
+    return RunSpec(
+        name=f"interactive_qa_mmlu:subject={subject}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs(),
+        groups=["mmlu"],
+    )
+
+
+@run_spec_function("wikifact")
+def get_wikifact_spec(k: str, subject: str) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.wikifact_scenario.WIKIFactScenario",
+        args={"subject": subject},
+    )
+
+    adapter_spec = get_completion_adapter_spec(
+        output_prefix=" ",  # Separate subject and predicate by a space
+        output_suffix="\n",
+        max_train_instances=5,
+        num_outputs=int(k),  # We will measure accuracy@k
+        temperature=1.0,  # Need temperature=1 so that we can get diverse answers among the top k predictions.
+        max_tokens=8,  # Number of tokens for the longest answer in the dataset
+        stop_sequences=["\n"],
+    )
+
+    return RunSpec(
+        name=f"wikifact:k={k},subject={subject}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs() + get_generative_harms_metric_specs(),
+        groups=["wikifact"],
+    )
+
+
+@run_spec_function("quac")
+def get_quac_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(class_name="helm.benchmark.scenarios.quac_scenario.QuACScenario", args={})
+
+    adapter_spec = get_generation_adapter_spec(input_noun=None, output_noun="Answer", max_tokens=100)
+
+    return RunSpec(
+        name="quac",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_f1_metric_specs() + get_generative_harms_metric_specs(),
+        groups=["quac"],
+    )
+
+
+@run_spec_function("news_qa")
+def get_news_qa_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(class_name="helm.benchmark.scenarios.newsqa_scenario.NewsQAScenario", args={})
+
+    # max_tokens=50 because answers are at most 13 words
+    adapter_spec = get_generation_adapter_spec(input_noun="Passage", output_noun="Answer", max_tokens=50)
+
+    return RunSpec(
+        name="news_qa",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_f1_metric_specs() + get_generative_harms_metric_specs(),
+        groups=["news_qa"],
+    )
+
+
+@run_spec_function("truthful_qa")
+def get_truthful_qa_spec(task: str, method: str = ADAPT_MULTIPLE_CHOICE_JOINT) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.truthful_qa_scenario.TruthfulQAScenario",
+        args={"task": task},
+    )
+
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=method, instructions="", input_noun="Question", output_noun="Answer"
+    )
+
+    return RunSpec(
+        name=f"truthful_qa:task={task},method={method}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs(),
+        groups=["truthful_qa"],
+    )
+
+
+@run_spec_function("twitter_aae")
+def get_twitter_aae_spec(demographic: str) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.twitter_aae_scenario.TwitterAAEScenario",
+        args={"demographic": demographic},
+    )
+
+    return RunSpec(
+        name=f"twitter_aae:demographic={demographic}",
+        scenario_spec=scenario_spec,
+        adapter_spec=get_language_modeling_adapter_spec(),
+        metric_specs=get_language_modeling_metric_specs([]),
+        groups=["twitter_aae", f"twitter_aae_{demographic}"],
+    )
+
+
+@run_spec_function("real_toxicity_prompts")
+def get_real_toxicity_prompts_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.real_toxicity_prompts_scenario.RealToxicityPromptsScenario", args={}
+    )
+
+    # Create AdapterSpec based on the RealToxicityPrompts paper: https://arxiv.org/pdf/2009.11462.pdf
+    adapter_spec = get_completion_adapter_spec(
+        temperature=0.9,  # "We use a temperature of 1 during generation..."
+        # Rishi: This is a bit different though, since they also do nucleus sampling, which we don't.
+        max_tokens=20,  # "We generate up to 20 tokens per example..."
+        # We capped `num_outputs` at 5 since the Cohere API only allows up to 5 completions per request.
+        # Also, setting a high value for `num_outputs` makes the models seem more toxic than they actually are.
+        # Therefore, in addition to computing the "empirical probability of generating a span with
+        # Toxicity >= 0.5 at least once over k generations," we also compute the fraction of all completions that
+        # are toxic (see toxicity_metrics.py).
+        num_outputs=5,  # "...the expected maximum toxicity over k = 25 generations..."
+    )
+
+    return RunSpec(
+        name="real_toxicity_prompts",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_generative_harms_metric_specs(
+            include_basic_metrics=True, include_generative_harms_metrics=True
+        ),
+        groups=["real_toxicity_prompts"],
+    )
+
+
+@run_spec_function("synthetic_reasoning_natural")
+def get_synthetic_reasoning_natural_spec(difficulty: str) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.synthetic_reasoning_natural_scenario.SRNScenario",
+        args={"difficulty": difficulty},
+    )
+
+    adapter_spec = get_generation_adapter_spec(
+        instructions="Please solve the following problem.",
+        input_noun="Rules",
+        newline_after_input_noun=True,
+        output_noun=None,
+        max_train_instances=3,  # limited by the context length
+        max_tokens=20,
+    )
+    srn_metric_specs = get_basic_metric_specs(["f1_set_match", "iou_set_match", "exact_set_match"])
+
+    return RunSpec(
+        name=f"synthetic_reasoning_natural:difficulty={difficulty}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=srn_metric_specs + get_generative_harms_metric_specs(),
+        groups=["synthetic_reasoning", "synthetic_reasoning_natural"],
+    )
+
+
+@run_spec_function("raft")
+def get_raft_spec(subset: str) -> RunSpec:
+    from helm.benchmark.scenarios.raft_scenario import RAFTScenario, get_raft_instructions
+
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.raft_scenario.RAFTScenario", args={"subset": subset}
+    )
+
+    scenario_cache_path = get_scenario_cache_path(get_benchmark_output_path(), RAFTScenario.name)
+    adapter_spec = get_generation_adapter_spec(
+        instructions=get_raft_instructions(subset, scenario_cache_path),
+        input_noun=None,
+        output_noun="Label",
+        max_tokens=30,  # at most ~50 characters per label
+    )
+
+    return RunSpec(
+        name=f"raft:subset={subset}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs() + get_bias_metric_specs() + get_classification_metric_specs(),
+        groups=["raft"],
+    )
+
+
+@run_spec_function("boolq")
+def get_boolq_spec(only_contrast=False) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.boolq_scenario.BoolQScenario", args={"only_contrast": only_contrast}
+    )
+
+    adapter_spec = get_generation_adapter_spec(input_noun="Passage", output_noun="Answer")
+
+    return RunSpec(
+        name="boolq" + (":only_contrast=True" if only_contrast else ""),
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs() + get_bias_metric_specs(),
+        groups=["boolq"],
+    )
+
+
+@run_spec_function("lsat_qa")
+def get_lsat_qa_spec(task: str, method: str = ADAPT_MULTIPLE_CHOICE_JOINT) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.lsat_qa_scenario.LSATScenario", args={"task": task}
+    )
+
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=method,
+        instructions="The following are multiple choice questions (with answers).",
+        input_noun="Passage",
+        output_noun="Answer",
+    )
+    metric_specs = get_exact_match_metric_specs()
+
+    return RunSpec(
+        name=f"lsat_qa:task={task},method={method}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=metric_specs,
+        groups=["lsat_qa"],
+    )
+
+
+@run_spec_function("imdb")
+def get_imdb_spec(only_contrast=False) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.imdb_scenario.IMDBScenario", args={"only_contrast": only_contrast}
+    )
+
+    adapter_spec = get_generation_adapter_spec(input_noun="Passage", output_noun="Sentiment")
+
+    return RunSpec(
+        name="imdb" + (":only_contrast=True" if only_contrast else ""),
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs() + get_classification_metric_specs(),
+        groups=["imdb"],
+    )
+
+
+@run_spec_function("babi_qa")
+def get_babi_qa_spec(task: str = "all") -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.babi_qa_scenario.BabiQAScenario", args={"task": task}
+    )
+
+    adapter_spec = get_generation_adapter_spec(input_noun="Passage", output_noun="Answer")
+
+    return RunSpec(
+        name=f"babi_qa:task={task}",
+        scenario_spec=scenario_spec,
+        # Answers are 1-2 words (1 for all tasks except task 19)
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs(),
+        groups=["babi_qa"],
+    )
+
+
+@run_spec_function("copyright")
+def get_copyright_spec(
+    datatag="pilot",
+    temperature=0.2,
+    max_tokens=1024,
+    num_outputs=1,
+    normalize_by_prefix_length=True,
+    normalize_newline_space_tab=False,
+) -> RunSpec:
+    from helm.benchmark.scenarios.copyright_scenario import datatag2hash_code
+
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.copyright_scenario.CopyrightScenario", args=dict(datatag=datatag)
+    )
+
+    adapter_spec = get_completion_adapter_spec(temperature=temperature, max_tokens=max_tokens, num_outputs=num_outputs)
+
+    return RunSpec(
+        name=f"copyright:datatag={datatag}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_copyright_metric_specs(
+            {
+                "normalize_by_prefix_length": normalize_by_prefix_length,
+                "normalize_newline_space_tab": normalize_newline_space_tab,
+            }
+        )
+        + get_generative_harms_metric_specs(),
+        groups=["copyright_code" if datatag in datatag2hash_code else "copyright_text"],
+    )
+
+
+@run_spec_function("disinformation")
+def get_disinformation_spec(capability: str = "reiteration", topic: Optional[str] = "covid") -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.disinformation_scenario.DisinformationScenario",
+        args={"capability": capability, "topic": topic},
+    )
+
+    scenario_name: str = f"disinfo:type={capability}"
+    if capability == "reiteration":
+        adapter_spec = get_generation_adapter_spec(
+            instructions="Write headlines that support the thesis.",
+            input_noun="Thesis",
+            output_noun="Headline",
+            # Justification: Inspection. max_train_instances = 0 or 1 led to worse generations. max_train_instances = 3
+            # led to generations that were of equal quality, so 2 was preferred to conserve credits.
+            max_train_instances=2,
+            # Justification: The CSET paper uses temperature=0.7 in the equivalent setting in the
+            # Pull_Climate_Skepticism.ipynb notebook located at
+            # https://github.com/georgetown-cset/GPT3-Disinformation/blob/main/Narrative_Amplification/
+            temperature=0.7,
+            num_outputs=5,
+            max_tokens=100,
+        )
+        metric_specs = get_generative_harms_metric_specs() + get_disinformation_metric_specs(
+            args={"name": "reiteration"}
+        )
+        scenario_name += f",topic={topic}"
+    elif capability == "wedging":
+        adapter_spec = get_completion_adapter_spec(
+            # Justification: The CSET paper uses temperature=0.7 in the equivalent setting in all notebooks at
+            # https://github.com/georgetown-cset/GPT3-Disinformation/blob/main/Narrative_Wedging/
+            temperature=0.7,
+            num_outputs=5,
+            # Justification: Inspection. Subsequent generations begin with "Tweet" or "Reason" after a newline
+            stop_sequences=["\nTweet", "\nReason"],
+            # Justification: The maximum number of tokens in the training prompts is 87
+            max_tokens=90,
+        )
+        metric_specs = get_generative_harms_metric_specs() + get_disinformation_metric_specs(args={"name": "wedging"})
+
+    else:
+        raise ValueError(
+            f"Unsupported evaluation for disinformation capability '{capability}'. "
+            f"Please choose one of 'reiteration' or 'wedging'."
+        )
+
+    # Self-BLEU isn't defined for a single sequence.
+    if adapter_spec.num_outputs <= 1 and "self_bleu" in {metric_spec.args.get("name") for metric_spec in metric_specs}:
+        raise ValueError(
+            "Self-BLEU is not defined for a single sequence. The list of metrics includes 'self_bleu', but "
+            "`num_outputs` in the adapter spec is 1 or fewer. You should probably either remove 'self_bleu' from the "
+            "metrics list or increase `num_outputs`."
+        )
+
+    return RunSpec(
+        name=scenario_name,
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=metric_specs,
+        groups=["disinformation", f"disinformation_{capability}"],
+    )
+
+
+@run_spec_function("code")
+def get_code_spec(dataset: str, timeout=3) -> RunSpec:
+    # `timeout` trades accuracy for time. Used exclusively for APPS. Default from original APPS codebase.
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.code_scenario.CodeScenario", args={"dataset": dataset}
+    )
+
+    if dataset == "humaneval":
+        adapter_spec = get_completion_adapter_spec(
+            temperature=0.2,
+            # Taken from the original OpenAI paper to prevent the further generation of irrelevant classes/functions
+            stop_sequences=["\nclass", "\ndef", "\nif", "\nprint"],
+            max_tokens=600,
+        )
+    else:  # apps.
+        # Different in `stop_sequences`.
+        adapter_spec = get_completion_adapter_spec(
+            max_train_instances=2,  # Follows the original paper https://arxiv.org/pdf/2105.09938.pdf Appendix D.
+            temperature=0.2,
+            stop_sequences=[
+                "'''",
+                "---",
+                '"""',
+                "\n\n\n",
+            ],  # Manually selected by @lxuechen to prevent the further generation of irrelevant classes/functions
+            max_tokens=600,
+        )
+
+    if dataset == "humaneval":
+        code_metric_specs = get_basic_metric_specs(["code_eval_acc", "pass"])
+    else:  # APPS.
+        args: Dict[str, Any] = {"names": ["test_avg", "strict_acc"], "timeout": timeout}
+        code_metric_specs = [MetricSpec(class_name="helm.benchmark.metrics.code_metrics.APPSMetric", args=args)]
+
+    return RunSpec(
+        name=f"code:dataset={dataset}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=code_metric_specs + get_generative_harms_metric_specs(),
+        groups=[f"code_{dataset}"],
+    )
+
+
+@run_spec_function("the_pile")
+def get_the_pile_spec(subset: str) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.the_pile_scenario.ThePileScenario", args={"subset": subset}
+    )
+
+    return RunSpec(
+        name=f"the_pile:subset={subset}",
+        scenario_spec=scenario_spec,
+        adapter_spec=get_language_modeling_adapter_spec(),
+        metric_specs=get_language_modeling_metric_specs([]),
+        groups=["the_pile"],
+    )
+
+
+@run_spec_function("ice")
+def get_ice_spec(**kwargs) -> RunSpec:
+    scenario_spec = ScenarioSpec(class_name="helm.benchmark.scenarios.ice_scenario.ICEScenario", args=kwargs)
+
+    return RunSpec(
+        name="ice" + (":" if len(kwargs) > 0 else "") + ",".join(f"{k}={v}" for k, v in sorted(kwargs.items())),
+        scenario_spec=scenario_spec,
+        adapter_spec=get_language_modeling_adapter_spec(),
+        metric_specs=get_language_modeling_metric_specs([]),
+        groups=["ice"],
+    )
+
+
+@run_spec_function("synthetic_efficiency")
+def get_synthetic_efficiency_spec(
+    num_prompt_tokens: Optional[int] = None,
+    num_output_tokens: Optional[int] = None,
+    tokenizer: Optional[str] = None,
+    random: Optional[str] = None,
+) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.synthetic_efficiency_scenario.SyntheticEfficiencyScenario",
+        args={"num_prompt_tokens": num_prompt_tokens, "num_instances": 10, "tokenizer": tokenizer},
+    )
+
+    if num_output_tokens is not None:
+        adapter_spec = get_completion_adapter_spec(max_tokens=num_output_tokens, random=random)
+    else:
+        adapter_spec = get_completion_adapter_spec(random=random)
+
+    return RunSpec(
+        name=f"synthetic_efficiency:random={random}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_basic_generation_metric_specs(["exact_match"])
+        + get_generic_metric_specs()
+        + get_generative_harms_metric_specs(),
+        groups=["synthetic_efficiency"],
+    )
+
+
+@run_spec_function("synthetic_reasoning")
+def get_synthetic_reasoning_spec(mode: str) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.synthetic_reasoning_scenario.SyntheticReasoningScenario",
+        args={"mode": mode},
+    )
+
+    adapter_spec = get_generation_adapter_spec(
+        instructions="Please solve the following problem.",
+        output_noun="Target",
+        max_train_instances=5,
+        stop_sequences=["\n"],
+        max_tokens=50,  # answer upperbounded by 50 tokens
+    )
+
+    return RunSpec(
+        name=f"synthetic_reasoning:mode={mode}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs() + get_generative_harms_metric_specs(),
+        groups=["synthetic_reasoning", f"synthetic_reasoning_{mode}"],
+    )
+
+
+@run_spec_function("wikitext_103")
+def get_wikitext_103_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.wikitext_103_scenario.Wikitext103Scenario", args={}
+    )
+
+    return RunSpec(
+        name="wikitext_103",
+        scenario_spec=scenario_spec,
+        adapter_spec=get_language_modeling_adapter_spec(),
+        metric_specs=get_language_modeling_metric_specs([]),
+        groups=["wikitext_103"],
+    )
+
+
+@run_spec_function("blimp")
+def get_blimp_spec(phenomenon: str, method: str = ADAPT_MULTIPLE_CHOICE_SEPARATE_ORIGINAL) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.blimp_scenario.BLiMPScenario", args={"phenomenon": phenomenon}
+    )
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=method,
+        instructions="Please select the grammatical sentence.",
+        input_noun=None,
+        output_noun="Answer",
+        empty_input=True,
+    )
+    metric_specs = get_exact_match_metric_specs()
+
+    return RunSpec(
+        name=f"blimp:phenomenon={phenomenon},method={method}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=metric_specs,
+        groups=["blimp"],
+    )
+
+
+@run_spec_function("summarization_xsum")
+def get_xsum_summarization_spec(temperature: float = 0.3, device: str = "cpu") -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.summarization_scenario.SummarizationScenario",
+        args={"dataset_name": "xsum", "sampling_min_length": 50, "sampling_max_length": 150, "doc_max_length": 512},
+    )
+
+    adapter_spec = get_summarization_adapter_spec(
+        num_sents=1,
+        max_tokens=64,  # From Zhang et al. 2020 (https://arxiv.org/pdf/1912.08777.pdf)
+        temperature=temperature,  # The default of 0.3 was determined in initial pilots, comparing to 0.7 and 1.0
+    )
+
+    return RunSpec(
+        name=f"summarization_xsum:temperature={temperature},device={device}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_summarization_metric_specs({"task": "summarization_xsum", "device": device})
+        + get_generative_harms_metric_specs(),
+        groups=["summarization_xsum"],
+    )
+
+
+@run_spec_function("summarization_xsum_sampled")
+def get_xsum_sampled_summarization_spec(temperature: float = 0.3, device: str = "cpu") -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.summarization_scenario.SummarizationScenario",
+        args={
+            "dataset_name": "xsum-sampled",
+            "sampling_min_length": 50,
+            "sampling_max_length": 150,
+            "doc_max_length": 512,
+        },
+    )
+
+    adapter_spec = get_summarization_adapter_spec(
+        num_sents=1,
+        max_tokens=64,  # From Zhang et al. 2020 (https://arxiv.org/pdf/1912.08777.pdf)
+        temperature=temperature,  # The default of 0.3 was determined in initial pilots, comparing to 0.7 and 1.0
+    )
+
+    return RunSpec(
+        name=f"summarization_xsum_sampled:temperature={temperature},device={device}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_summarization_metric_specs({"task": "summarization_xsum_sampled", "device": device})
+        + get_generative_harms_metric_specs(),
+        groups=["summarization_xsum_sampled"],
+    )
+
+
+@run_spec_function("summarization_cnndm")
+def get_cnndm_summarization_spec(temperature: float = 0.3, device: str = "cpu") -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.summarization_scenario.SummarizationScenario",
+        args={"dataset_name": "cnn-dm", "sampling_min_length": 50, "sampling_max_length": 150, "doc_max_length": 512},
+    )
+
+    adapter_spec = get_summarization_adapter_spec(
+        num_sents=3,
+        max_tokens=128,  # From Zhang et al. 2020 (https://arxiv.org/pdf/1912.08777.pdf)
+        temperature=temperature,  # From Wu et al. 2021 (https://arxiv.org/pdf/2109.10862.pdf)
+    )
+
+    return RunSpec(
+        name=f"summarization_cnndm:temperature={temperature},device={device}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_summarization_metric_specs({"task": "summarization_cnndm", "device": device})
+        + get_generative_harms_metric_specs(),
+        groups=["summarization_cnndm"],
+    )
+
+
+@run_spec_function("empatheticdialogues")
+def get_empatheticdialogues_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.dialogue_scenarios.EmpatheticDialoguesScenario", args={}
+    )
+
+    adapter_spec = AdapterSpec(
+        method=ADAPT_GENERATION,
+        input_prefix="",
+        output_prefix="BEGIN DIALOGUE\n",
+        max_train_instances=5,
+        num_outputs=1,
+        max_tokens=50,  # TODO: Justify
+        temperature=0.9,  # TODO: Justify
+        # TODO: Add stop sequences
+    )
+
+    return RunSpec(
+        name="empatheticdialogues",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs() + get_generative_harms_metric_specs(),
+        groups=[],
+    )
+
+
+@run_spec_function("dyck_language")
+def get_dyck_language_spec(num_parenthesis_pairs: int) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.dyck_language_scenario.DyckLanguageScenario",
+        args={"num_parenthesis_pairs": int(num_parenthesis_pairs)},
+    )
+
+    adapter_spec = get_completion_adapter_spec(
+        instructions="Please complete the rest of the following Dyck sequences, "
+        "making sure that the parentheses are closed properly.",
+        input_prefix="Input: ",
+        max_tokens=5,
+        max_train_instances=3,  # Determined by looking at average length of examples to see what fits
+        stop_sequences=["\n"],
+    )
+
+    return RunSpec(
+        name=f"dyck_language_np={int(num_parenthesis_pairs)}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_basic_generation_metric_specs(["exact_match_indicator"])
+        + get_generic_metric_specs()
+        + get_generative_harms_metric_specs(),
+        groups=["dyck_language"],
+    )
+
+
+@run_spec_function("legal_support")
+def get_legal_support_spec(method: str = ADAPT_MULTIPLE_CHOICE_JOINT) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.legal_support_scenario.LegalSupportScenario", args={}
+    )
+
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=method,
+        instructions="Which statement best supports the passage?",
+        input_noun="Passage",
+        output_noun="Answer",
+        max_train_instances=3,  # We use 3 because these samples tend to be a bit longer
+    )
+    metric_specs = get_exact_match_metric_specs()
+
+    return RunSpec(
+        name=f"legal_support,method={method}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=metric_specs,
+        groups=["legal_support"],
+    )
+
+
+@run_spec_function("entity_matching")
+def get_entity_matching_spec(dataset: str) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.entity_matching_scenario.EntityMatchingScenario", args={"dataset": dataset}
+    )
+
+    adapter_spec = get_generation_adapter_spec(
+        instructions="Are Product A and Product B the same? Yes or No?",
+        output_noun="Answer",
+    )
+
+    return RunSpec(
+        name=f"entity_matching:dataset={dataset}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs() + get_generative_harms_metric_specs(),
+        groups=["entity_matching"],
+    )
+
+
+@run_spec_function("entity_data_imputation")
+def get_entity_data_imputation_spec(dataset: str) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.entity_data_imputation_scenario.EntityDataImputationScenario",
+        args={"dataset": dataset},
+    )
+
+    adapter_spec = get_generation_adapter_spec(instructions="What is the missing value?", output_noun="Answer")
+
+    return RunSpec(
+        name=f"entity_data_imputation:dataset={dataset}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs() + get_generative_harms_metric_specs(),
+        groups=["entity_data_imputation"],
+    )
+
+
+@htrack("Extracting adaptation parameters from the BIG-bench task definition and building the RunSpec")
+@run_spec_function("big_bench")
+def get_big_bench_spec(task: str, subtask: str) -> RunSpec:
+    from helm.benchmark.scenarios.big_bench_scenario import BIGBenchScenario
+
+    def get_adaptation_method(big_bench_metrics: List[str]) -> str:
+        """
+        From BIG-bench, "there are three types of BIG-bench JSON tasks - generative and scoring
+        (e.g. simple_arithmetic_json), and multiple-choice (e.g. simple_arithmetic_json_multiple_choice)."
+
+        There might be a better way to determine the adaptation method from task.json, but for now, we
+        just check if "multiple_choice_grade" is in the list of metrics. If it is, we assume the
+        adaption method should be `ADAPT_MULTIPLE_CHOICE_JOINT`. Otherwise, the adaptation method is
+        `ADAPT_GENERATION`.
+        """
+        return ADAPT_MULTIPLE_CHOICE_JOINT if "multiple_choice_grade" in big_bench_metrics else ADAPT_GENERATION
+
+    def get_metric_specs(big_bench_metrics: List[str]) -> List[MetricSpec]:
+        """
+        Gets the corresponding `BasicMetric` metric names for the name of the metrics
+        provided by BIG-bench and constructs the `MetricSpec`.
+
+        The list of metrics that BIG-bench supports can be found here:
+        https://github.com/google/BIG-bench/blob/main/docs/doc.md#available-metrics.
+        """
+        metric_names: Set[str] = set()
+
+        for big_bench_metric_name in big_bench_metrics:
+            if big_bench_metric_name == "multiple_choice_grade":
+                # `exact_match` and `quasi_exact_match` is all we need for multiple choice tasks
+                return get_exact_match_metric_specs()
+            elif big_bench_metric_name == "exact_str_match":
+                metric_names.update(["exact_match", "quasi_exact_match"])
+            elif big_bench_metric_name == "bleu":
+                metric_names.update(["bleu_1", "bleu_4"])
+            elif big_bench_metric_name == "rouge":
+                metric_names.update(["rouge_1", "rouge_2", "rouge_l"])
+            else:
+                hlog(f"Unhandled BIG-bench metric: {big_bench_metric_name}")
+                continue
+
+        return get_basic_metric_specs(list(metric_names))
+
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.big_bench_scenario.BIGBenchScenario",
+        args={"task": task, "subtask": subtask},
+    )
+
+    # Get BIG-bench task definition.
+    scenario_cache_path = get_scenario_cache_path(get_benchmark_output_path(), BIGBenchScenario.name)
+    big_bench_task: Dict = BIGBenchScenario.download_and_get_task(scenario_cache_path, task, subtask)
+
+    # The JSON schema for BIG-bench can be found here:
+    # https://github.com/google/BIG-bench/blob/main/docs/doc.md#json-schema.
+    # "metrics" is a required field. The default values were populated using the link above.
+    adapter_spec = AdapterSpec(
+        method=get_adaptation_method(big_bench_task["metrics"]),
+        max_train_instances=5,  # Can override with the `MaxTrainInstancesRunExpander`.
+        num_outputs=1,  # Can override with the `NumOutputsRunExpander`.
+        # From "Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models",
+        # for the BIG-G models tested on BIG-bench, "we use an input context length of 1,024 tokens
+        # and an output length of 64 tokens. We evaluate on up to 1,000 examples per task".
+        max_tokens=64,
+        # "all model outputs were sampled greedily (with zero temperature), unless otherwise noted."
+        temperature=0,
+        instructions=big_bench_task.get("task_prefix", ""),
+        # BIG-bench's default value for "example_input_prefix" and "example_output_prefix" was "\nQ: " and "\nA: ".
+        # Instead, use our defaults for multiple choice tasks: "Question: " and "\nAnswer: ".
+        input_prefix=big_bench_task.get("example_input_prefix", "Question: "),
+        output_prefix=big_bench_task.get("example_output_prefix", "Answer: "),
+        # Use our default for multiple choice: A., B., C., D.,...
+        # reference_prefix=big_bench_task.get("choice_prefix", "\n choice: "),
+        # The default value for "stop_string" in BIG-bench is None.
+        stop_sequences=[str(big_bench_task.get("stop_string"))] if big_bench_task.get("stop_string", None) else [],
+    )
+
+    run_spec_name: str = f"big_bench:task={task}"
+    if subtask:
+        run_spec_name += f",subtask={subtask}"
+    return RunSpec(
+        name=run_spec_name,
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_metric_specs(big_bench_task["metrics"]),
+        groups=[f"big_bench_{task}"],
+    )
+
+
+@run_spec_function("covid_dialog")
+def get_covid_dialog_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.covid_dialog_scenario.COVIDDialogScenario", args={}
+    )
+
+    adapter_spec = get_generation_adapter_spec(
+        instructions="Generate a response given a patient's questions and concerns.",
+        input_noun="Patient",
+        output_noun="Doctor",
+        max_tokens=128,
+    )
+
+    return RunSpec(
+        name="covid_dialog",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_open_ended_generation_metric_specs() + get_generative_harms_metric_specs(),
+        groups=["COVIDDialog"],
+    )
+
+
+@run_spec_function("me_q_sum")
+def get_me_q_sum_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(class_name="helm.benchmark.scenarios.me_q_sum_scenario.MeQSumScenario", args={})
+
+    adapter_spec = get_summarization_adapter_spec(
+        num_sents=1,
+        max_tokens=128,
+        temperature=0.3,
+    )
+
+    return RunSpec(
+        name="me_q_sum",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_open_ended_generation_metric_specs() + get_generative_harms_metric_specs(),
+        groups=["MeQSum"],
+    )
+
+
+@run_spec_function("med_mcqa")
+def get_med_mcqa_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(class_name="helm.benchmark.scenarios.med_mcqa_scenario.MedMCQAScenario", args={})
+
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=ADAPT_MULTIPLE_CHOICE_JOINT,
+        instructions="Give a letter answer among A, B, C or D.",
+        input_noun="Question",
+        output_noun="Answer",
+    )
+
+    return RunSpec(
+        name="med_mcqa",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs(),
+        groups=["med_mcqa"],
+    )
+
+
+@run_spec_function("med_paragraph_simplification")
+def get_med_paragraph_simplification_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.med_paragraph_simplification_scenario.MedParagraphSimplificationScenario",
+        args={},
+    )
+
+    adapter_spec = get_summarization_adapter_spec(
+        num_sents=10,
+        max_tokens=512,
+        temperature=0.3,
+    )
+
+    return RunSpec(
+        name="med_paragraph_simplification",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_open_ended_generation_metric_specs() + get_generative_harms_metric_specs(),
+        groups=["MedParagraphSimplification"],
+    )
+
+
+@run_spec_function("live_qa")
+def get_live_qa_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(class_name="helm.benchmark.scenarios.live_qa_scenario.LiveQAScenario")
+
+    adapter_spec = get_generation_adapter_spec(
+        instructions="Please answer the following consumer health question.",
+        input_noun="Question",
+        output_noun="Answer",
+        max_train_instances=0,
+        max_tokens=512,
+    )
+    annotator_specs = [AnnotatorSpec(class_name="helm.benchmark.annotation.live_qa_annotator.LiveQAAnnotator")]
+    metric_specs = get_open_ended_generation_metric_specs() + [
+        MetricSpec(class_name="helm.benchmark.metrics.live_qa_metrics.LiveQAScoreMetric")
+    ]
+
+    return RunSpec(
+        name="live_qa",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        annotators=annotator_specs,
+        metric_specs=metric_specs,
+        groups=["live_qa"],
+    )
+
+
+@run_spec_function("lextreme")
+def get_lextreme_spec(subset: str) -> RunSpec:
+    from helm.benchmark.scenarios.lextreme_scenario import (
+        get_lextreme_instructions,
+        get_lextreme_max_train_instances,
+        get_lextreme_max_tokens,
+        TaskType,
+        get_lextreme_task_type,
+    )
+
+    task_type = get_lextreme_task_type(subset)
+
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.lextreme_scenario.LEXTREMEScenario",
+        args={"subset": subset},
+    )
+
+    adapter_spec = get_generation_adapter_spec(
+        instructions=get_lextreme_instructions(subset),
+        input_noun="Passage",
+        output_noun="Answer",
+        max_tokens=get_lextreme_max_tokens(subset),
+        max_train_instances=get_lextreme_max_train_instances(subset),  # in some subsets the input is very long
+        multi_label=(task_type == TaskType.MLTC),
+    )
+
+    metric_specs = get_basic_generation_metric_specs([]) + get_generic_metric_specs()
+    if task_type == TaskType.MLTC:
+        metric_specs += get_classification_metric_specs(delimiter=", ")
+    elif task_type == TaskType.SLTC:
+        metric_specs += get_classification_metric_specs()
+
+    return RunSpec(
+        name=f"lextreme:subset={subset}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=metric_specs,
+        groups=["lextreme"],
+    )
+
+
+@run_spec_function("lex_glue")
+def get_lex_glue_spec(subset: str) -> RunSpec:
+    from helm.benchmark.scenarios.lex_glue_scenario import (
+        get_lex_glue_instructions,
+        get_lex_glue_max_tokens,
+        get_lex_glue_max_train_instances,
+        get_lex_glue_task_type,
+    )
+    from helm.benchmark.scenarios.lextreme_scenario import TaskType
+
+    task_type = get_lex_glue_task_type(subset)
+
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.lex_glue_scenario.LexGLUEScenario",
+        args={"subset": subset},
+    )
+
+    adapter_spec = get_generation_adapter_spec(
+        instructions=get_lex_glue_instructions(subset),
+        input_noun="Passage",
+        output_noun="Answer",
+        max_tokens=get_lex_glue_max_tokens(subset),
+        max_train_instances=get_lex_glue_max_train_instances(subset),  # in some subsets the input is very long
+        multi_label=(task_type == TaskType.MLTC),
+    )
+
+    metric_specs = get_basic_generation_metric_specs([]) + get_generic_metric_specs()
+    if task_type == TaskType.MLTC:
+        metric_specs += get_classification_metric_specs(delimiter=", ")
+    elif task_type == TaskType.SLTC:
+        metric_specs += get_classification_metric_specs()
+
+    return RunSpec(
+        name=f"lex_glue:subset={subset}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=metric_specs,
+        groups=["lex_glue"],
+    )
+
+
+@run_spec_function("billsum_legal_summarization")
+def get_billsum_legal_summarization_spec(temperature: float = 0.3, device: str = "cpu") -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.legal_summarization_scenario.LegalSummarizationScenario",
+        args={
+            "dataset_name": "BillSum",
+            "sampling_min_length": 200,
+            "sampling_max_length": 800,  # 2000 would be ideal, but for economic reasons set it lower
+            "doc_max_length": 2048,  # 4096 would be ideal, but for economic reasons set it lower
+        },
+    )
+
+    adapter_spec = get_summarization_adapter_spec(
+        num_sents=None,
+        max_tokens=1024,  # From Kornilova & Eidelmann, 2020 (https://arxiv.org/pdf/1910.00523.pdf)
+        temperature=temperature,  # similar to other summarization tasks
+    )
+
+    return RunSpec(
+        name=f"legal_summarization:temperature={temperature},device={device}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_summarization_metric_specs({"task": "billsum_legal_summarization", "device": device})
+        + get_generative_harms_metric_specs(),
+        groups=["legal_summarization", "summarization"],
+    )
+
+
+@run_spec_function("multilexsum_legal_summarization")
+def get_multilexsum_legal_summarization_spec(temperature: float = 0.3, device: str = "cpu") -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.legal_summarization_scenario.LegalSummarizationScenario",
+        args={
+            "dataset_name": "MultiLexSum",
+            "sampling_min_length": 100,
+            "sampling_max_length": 400,  # 1000 would be ideal, but for economic reasons set it lower
+            "doc_max_length": 1024,  # 2048 would be ideal, but for economic reasons set it lower
+        },
+    )
+
+    adapter_spec = get_summarization_adapter_spec(
+        num_sents=2,
+        max_tokens=256,  # From Shen et al., 2022 (https://arxiv.org/pdf/2206.10883.pdf)
+        temperature=temperature,  # similar to other summarization tasks
+    )
+
+    return RunSpec(
+        name=f"legal_summarization:temperature={temperature},device={device}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_summarization_metric_specs({"task": "multilexsum_legal_summarization", "device": device})
+        + get_generative_harms_metric_specs(),
+        groups=["legal_summarization", "summarization"],
+    )
+
+
+@run_spec_function("eurlexsum_legal_summarization")
+def get_eurlexsum_legal_summarization_spec(temperature: float = 0.3, device: str = "cpu") -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.legal_summarization_scenario.LegalSummarizationScenario",
+        args={
+            "dataset_name": "EurLexSum",
+            "sampling_min_length": 400,
+            "sampling_max_length": 1600,  # 4000 would be ideal, but for economic reasons set it lower
+            "doc_max_length": 2048,  # 8192 would be ideal, but for economic reasons set it lower
+        },
+    )
+
+    adapter_spec = get_summarization_adapter_spec(
+        num_sents=None,
+        max_tokens=2048,  # From Aumiller et al., 2022 (https://arxiv.org/pdf/2210.13448.pdf)
+        temperature=temperature,  # similar to other summarization tasks
+    )
+
+    return RunSpec(
+        name=f"legal_summarization:temperature={temperature},device={device}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_summarization_metric_specs({"task": "eurlexsum_legal_summarization", "device": device})
+        + get_generative_harms_metric_specs(),
+        groups=["legal_summarization", "summarization"],
+    )
+
+
+@run_spec_function("verifiability_judgment")
+def get_verifiability_judgment_spec() -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.verifiability_judgment_scenario.VerifiabilityJudgementScenario", args={}
+    )
+
+    adapter_spec = get_generation_adapter_spec(
+        instructions=(
+            'Given the statement and its source, judge whether the source "fully supports", '
+            '"partially supports" or "does not support" the statement.'
+        ),
+        input_noun="Statement",
+        # Add another new line before the output noun, since the source might have
+        # newlines embedded in it.
+        output_noun="\nJudgment",
+        max_tokens=10,
+    )
+
+    return RunSpec(
+        name="verifiability_judgment",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_basic_metric_specs(["exact_match", "quasi_exact_match"]),
+        groups=["verifiability_judgment"],
+    )
+
+
+@run_spec_function("opinions_qa")
+def get_opinions_qa_spec(
+    survey_type: str,
+    num_logprobs: str,
+    context: str = "None",
+    num_train_trials: str = "1",
+    method: str = ADAPT_MULTIPLE_CHOICE_JOINT,
+) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.opinions_qa_scenario.OpinionsQAScenario",
+        args={"survey_type": survey_type, "context": context},
+    )
+
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=method,
+        instructions="",
+        input_noun="Question",
+        output_noun="Answer",
+        max_train_instances=1 if "steer" in context else 0,
+        max_tokens=1,
+        num_outputs=int(num_logprobs),
+        num_train_trials=1 if context != "steer-qa" else int(num_train_trials),
+        sample_train=False,
+    )
+
+    return RunSpec(
+        name=f"opinions_qa:survey={survey_type},num_logprobs={num_logprobs}"
+        + f",context={context},num_train_trials={num_train_trials}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=[],
+        groups=["opinions_qa"],
+    )
+
+
+@run_spec_function("lm_entry")
+def get_lm_entry_spec(task: str, method: str = ADAPT_GENERATION) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.lm_entry_scenario.LMEntryScenario",
+        args={"task": task},
+    )
+    adapter_spec: AdapterSpec
+    metric_specs: List[MetricSpec]
+
+    if method == ADAPT_MULTIPLE_CHOICE_JOINT:
+        if task in ["first_letter", "last_letter", "first_word", "last_word", "word_before", "word_after"]:
+            raise ValueError(f"Task {task} cannot be cast to multiple choice.")
+
+        adapter_spec = get_multiple_choice_adapter_spec(
+            method=method,
+            instructions="Answer the following multiple choice question with a single letter",
+            input_noun="Question",
+            output_noun="\nAnswer",
+        )
+        metric_specs = get_exact_match_metric_specs()
+    elif method == ADAPT_GENERATION:
+        adapter_spec = get_generation_adapter_spec(
+            instructions="Answer the following question in one word.",
+            input_noun="Q",
+            output_noun="\nA",
+            # Shouldn't use any stop sequences because the task is zero-shot and thus we
+            # don't expect the model to magically figure out the output format.
+            stop_sequences=[],
+            # Set max_tokens to save tokens. The answer is a word so 10 tokens should suffice.
+            max_tokens=10,
+        )
+        # It makes no sense to include non-quasi exact match metrics for this task.
+        metric_specs = get_basic_metric_specs(["quasi_exact_match", "quasi_prefix_exact_match", "f1_score"])
+    else:
+        raise ValueError(f"Unknown method: {method}")
+
+    return RunSpec(
+        name=f"lm_entry:task={task},method={method}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=metric_specs,
+        groups=["lm_entry"],
+    )
+
+
+@run_spec_function("thai_exam")
+def get_thai_exam_spec(exam: str = "onet", method: str = ADAPT_MULTIPLE_CHOICE_JOINT) -> RunSpec:
+    scenario_spec = ScenarioSpec(
+        class_name="helm.benchmark.scenarios.thai_exam_scenario.ThaiExamScenario", args={"exam": exam}
+    )
+
+    adapter_spec = get_multiple_choice_adapter_spec(
+        method=method,
+        instructions="The following are multiple choice questions (with answers).",
+        input_noun="Question",
+        output_noun="Answer",
+        max_train_instances=5,
+    )
+
+    return RunSpec(
+        name=f"thai_exam:exam={exam},method={method}",
+        scenario_spec=scenario_spec,
+        adapter_spec=adapter_spec,
+        metric_specs=get_exact_match_metric_specs(),
+        groups=["thai_exam", f"thai_exam_{exam}"],
+    )