crfm-helm 0.3.0__py3-none-any.whl → 0.5.0__py3-none-any.whl

helm/benchmark/augmentations/person_name_perturbation.py

@@ -263,13 +263,6 @@ class PersonNamePerturbation(Perturbation):
         name = rng.choice(list(options))
         return name
 
-    def perturb(self, text: str, rng: Random) -> str:
-        """
-        Perturbing the text is handled in `perturb_with_persistency` to ensure that perturbed names
-        in `Instance`s and `Reference`s match.
-        """
-        raise NotImplementedError("Should never be called")
-
     def perturb_with_persistency(
         self, text: str, rng: Random, name_substitution_mapping: Dict[str, str], skipped_tokens: Set[str]
     ) -> str:

helm/benchmark/augmentations/perturbation.py

@@ -10,7 +10,6 @@ from helm.common.object_spec import ObjectSpec, create_object
 
 
 class Perturbation(ABC):
-
     # Unique name to describe perturbation
     name: str
 
@@ -28,17 +27,24 @@ class Perturbation(ABC):
         # If seed exists, use it as part of the random seed
         return Random(instance.id if seed is None else str(seed) + instance.id)
 
+    @abstractmethod
+    def apply(self, instance: Instance, seed: Optional[int] = None) -> Instance:
+        """Generate a modified instance from the input instance."""
+        pass
+
+
+class TextPerturbation(Perturbation, ABC):
     def apply(self, instance: Instance, seed: Optional[int] = None) -> Instance:
         """
-        Generates a new Instance by perturbing the input, tagging the Instance and perturbing the References,
-        if should_perturb_references is true. Initializes a random number generator based on instance_id that gets
-        passed to perturb and perturb_references.
+        Generates a new Instance by applying `perturb` to the input and (if requested) the references.
+        Initializes a random number generator based on instance_id that gets
+        passed to perturb.
         """
         rng: Random = self.get_rng(instance, seed)
 
         references: List[Reference] = instance.references
         if self.should_perturb_references:
-            references = [self.perturb_reference(reference, rng) for reference in references]
+            references = [self._perturb_reference(reference, rng) for reference in references]
 
         description = replace(self.description, seed=seed)
 
@@ -49,11 +55,18 @@ class Perturbation(ABC):
             input=Input(text=self.perturb(instance.input.text, rng)),
             references=references,
             perturbation=description,
+            contrast_inputs=[instance.input],
         )
 
-    def perturb_reference(self, reference: Reference, rng: Random) -> Reference:
+    def _perturb_reference(self, reference: Reference, rng: Random) -> Reference:
         """Generates a new Reference by perturbing the output and tagging the Reference."""
-        return replace(reference, output=Output(text=self.perturb(reference.output.text, rng)), tags=reference.tags)
+        return replace(
+            reference,
+            output=Output(
+                text=self.perturb(reference.output.text, rng), multimedia_content=reference.output.multimedia_content
+            ),
+            tags=reference.tags,
+        )
 
     @abstractmethod
     def perturb(self, text: str, rng: Random) -> str:

helm/benchmark/augmentations/perturbation_description.py

@@ -23,7 +23,7 @@ class PerturbationDescription:
     computed_on: str = PERTURBATION_PERTURBED
     """Which types of Instances we are evaluating, to be populated during metric evaluation. PERTURBATION_PERTURBED
     (default) means we are evaluating on perturbed instances, PERTURBATION_ORIGINAL means we are evaluating the
-    unperturbed version of instances where this perturbation appplies, and, PERTURBATION_WORST means the the minimum
+    unperturbed version of instances where this perturbation applies, and, PERTURBATION_WORST means the the minimum
     metric between the two."""
 
     seed: Optional[int] = None

helm/benchmark/augmentations/space_perturbation.py

@@ -2,11 +2,11 @@ from dataclasses import dataclass
 from random import Random
 import re
 
-from .perturbation import Perturbation
+from .perturbation import TextPerturbation
 from .perturbation_description import PerturbationDescription
 
 
-class SpacePerturbation(Perturbation):
+class SpacePerturbation(TextPerturbation):
     """
     A simple perturbation that replaces existing spaces with 0-max_spaces spaces (thus potentially merging words).
     """

helm/benchmark/augmentations/suffix_perturbation.py

@@ -0,0 +1,29 @@
+from dataclasses import dataclass
+from random import Random
+
+from .perturbation import TextPerturbation
+from .perturbation_description import PerturbationDescription
+
+
+class SuffixPerturbation(TextPerturbation):
+    """
+    Appends a suffix to the end of the text. Example:
+
+    A picture of a dog -> A picture of a dog, picasso
+    """
+
+    @dataclass(frozen=True)
+    class Description(PerturbationDescription):
+        suffix: str = ""
+
+    name: str = "style"
+
+    def __init__(self, suffix: str):
+        self._suffix: str = suffix
+
+    @property
+    def description(self) -> PerturbationDescription:
+        return SuffixPerturbation.Description(name=self.name, suffix=self._suffix)
+
+    def perturb(self, text: str, rng: Random) -> str:
+        return f"{text}, {self._suffix}"

helm/benchmark/augmentations/synonym_perturbation.py

@@ -11,10 +11,10 @@ import spacy
 
 from helm.common.general import match_case, ensure_file_downloaded
 from .perturbation_description import PerturbationDescription
-from .perturbation import Perturbation
+from .perturbation import TextPerturbation
 
 
-class SynonymPerturbation(Perturbation):
+class SynonymPerturbation(TextPerturbation):
     """
     Synonyms. For implementation details, see
     https://github.com/GEM-benchmark/NL-Augmenter/blob/main/nlaugmenter/transformations/synonym_substitution/transformation.py

helm/benchmark/augmentations/test_perturbation.py

@@ -15,6 +15,7 @@ from .space_perturbation import SpacePerturbation
 from .dialect_perturbation import DialectPerturbation
 from .person_name_perturbation import PersonNamePerturbation
 from .gender_perturbation import GenderPerturbation
+from .suffix_perturbation import SuffixPerturbation
 
 
 def test_extra_space_perturbation():
@@ -145,7 +146,6 @@ def test_space_perturbation():
     instance: Instance = Instance(id="id0", input=Input(text="Hello World!\nQuite a day, huh?"), references=[])
     instances: List[Instance] = data_augmenter.generate([instance], include_original=True)
 
-    print(instances)
     assert len(instances) == 2
     assert instances[1].perturbation.name == "space"
     assert instances[1].input.text == "Hello   World!\nQuite a  day,   huh?"
@@ -162,7 +162,6 @@ def test_dialect_perturbation():
     )
     instances: List[Instance] = data_augmenter.generate([instance], include_original=True)
 
-    print(instances)
     assert len(instances) == 2
     assert instances[1].perturbation.name == "dialect"
     assert instances[1].input.text == "I gon remember dis day to b the best day of mah life."
@@ -188,7 +187,6 @@ def test_person_name_perturbation():
     )
     instances: List[Instance] = data_augmenter.generate([instance], include_original=True)
 
-    print(instances)
     assert len(instances) == 2
     assert instances[1].perturbation.name == "person_name"
     assert (
@@ -209,7 +207,6 @@ def test_gender_pronoun_perturbation():
     )
     instances: List[Instance] = data_augmenter.generate([instance], include_original=True)
 
-    print(instances)
     assert len(instances) == 2
     assert instances[1].perturbation.mode == "pronouns"
     assert instances[1].input.text == "Did she mention that she was coming with her parents and their friends?"
@@ -227,13 +224,22 @@ def test_gender_term_perturbation():
     )
     instances: List[Instance] = data_augmenter.generate([instance], include_original=True)
 
-    print(instances)
     assert len(instances) == 2
     assert instances[1].perturbation.mode == "terms"
     assert instances[1].input.text == "His granddaughters looked a lot like their mom."
     assert instances[1].references[0].output.text == "How did their mother look like?"
 
 
+def test_suffix_perturbation():
+    data_augmenter = DataAugmenter(perturbations=[SuffixPerturbation(suffix="pixel art")])
+    instance: Instance = Instance(id="id0", input=Input(text="A blue dog"), references=[])
+    instances: List[Instance] = data_augmenter.generate([instance], include_original=True)
+
+    assert len(instances) == 2
+    assert instances[1].perturbation.suffix == "pixel art"
+    assert instances[1].input.text == "A blue dog, pixel art"
+
+
 # TODO(#1958) Fix the logic to renable this test
 @unittest.skip("Currently cannot replace words at either text boundary.")
 def test_gender_term_perturbation_edge_word():
@@ -247,7 +253,6 @@ def test_gender_term_perturbation_edge_word():
     )
     instances: List[Instance] = data_augmenter.generate([instance], include_original=False)
 
-    print(instances)
     assert len(instances) == 1
     assert instances[0].input.text == "mom said it is okay"
     assert instances[0].references[0].output.text == "Sure he did daughter"
@@ -266,6 +271,5 @@ def test_gender_term_perturbation_consequtive_words():
     )
     instances: List[Instance] = data_augmenter.generate([instance], include_original=False)
 
-    print(instances)
     assert len(instances) == 1
     assert instances[0].input.text == "I'm a mom mom: my daughter has a daughter."

helm/benchmark/augmentations/translate_perturbation.py

@@ -0,0 +1,30 @@
+from dataclasses import dataclass
+from random import Random
+
+from helm.clients.google_translate_client import GoogleTranslateClient
+from .perturbation import TextPerturbation
+from .perturbation_description import PerturbationDescription
+
+
+class TranslatePerturbation(TextPerturbation):
+    """
+    Translates to different languages.
+    """
+
+    @dataclass(frozen=True)
+    class Description(PerturbationDescription):
+        # Language code to translate to. Needs a default value since we inherit from `PerturbationDescription`
+        language_code: str = "zh-CN"
+
+    name: str = "translate"
+
+    def __init__(self, language_code: str):
+        self.language_code: str = language_code
+        self.google_translate_client = GoogleTranslateClient()
+
+    @property
+    def description(self) -> PerturbationDescription:
+        return TranslatePerturbation.Description(name=self.name, language_code=self.language_code)
+
+    def perturb(self, text: str, rng: Random) -> str:
+        return self.google_translate_client.translate(text, self.language_code)

helm/benchmark/augmentations/typos_perturbation.py

@@ -2,10 +2,10 @@ from dataclasses import dataclass
 from random import Random
 
 from .perturbation_description import PerturbationDescription
-from .perturbation import Perturbation
+from .perturbation import TextPerturbation
 
 
-class TyposPerturbation(Perturbation):
+class TyposPerturbation(TextPerturbation):
     """
     Typos. For implementation details, see
     https://github.com/GEM-benchmark/NL-Augmenter/tree/main/transformations/butter_fingers_perturbation

helm/benchmark/config_registry.py

@@ -0,0 +1,38 @@
+import os
+import importlib_resources as resources
+
+from helm.benchmark.model_deployment_registry import register_model_deployments_from_path
+from helm.benchmark.model_metadata_registry import register_model_metadata_from_path
+from helm.benchmark.tokenizer_config_registry import register_tokenizer_configs_from_path
+from helm.benchmark.runner_config_registry import register_runner_config_from_path
+
+
+MODEL_METADATA_FILE: str = "model_metadata.yaml"
+TOKENIZER_CONFIGS_FILE: str = "tokenizer_configs.yaml"
+MODEL_DEPLOYMENTS_FILE: str = "model_deployments.yaml"
+RUNNER_CONFIG_FILE: str = "runner_config.yaml"
+
+CONFIG_PACKAGE = "helm.config"
+
+
+def register_configs_from_directory(dir_path: str) -> None:
+    model_metadata_path = os.path.join(dir_path, MODEL_METADATA_FILE)
+    if os.path.isfile(model_metadata_path):
+        register_model_metadata_from_path(model_metadata_path)
+
+    tokenizer_configs_path = os.path.join(dir_path, TOKENIZER_CONFIGS_FILE)
+    if os.path.isfile(tokenizer_configs_path):
+        register_tokenizer_configs_from_path(tokenizer_configs_path)
+
+    model_deployments_path = os.path.join(dir_path, MODEL_DEPLOYMENTS_FILE)
+    if os.path.isfile(model_deployments_path):
+        register_model_deployments_from_path(model_deployments_path)
+
+    runner_config_path = os.path.join(dir_path, RUNNER_CONFIG_FILE)
+    if os.path.isfile(runner_config_path):
+        register_runner_config_from_path(runner_config_path)
+
+
+def register_builtin_configs_from_helm_package() -> None:
+    package_path = str(resources.files(CONFIG_PACKAGE))
+    register_configs_from_directory(package_path)

helm/benchmark/executor.py

@@ -1,9 +1,15 @@
 from typing import Optional
 from dataclasses import dataclass, replace
+from helm.common.cache_backend_config import (
+    CacheBackendConfig,
+    BlackHoleCacheBackendConfig,
+    MongoCacheBackendConfig,
+    SqliteCacheBackendConfig,
+)
 
 from helm.common.general import parallel_map
 from helm.common.hierarchical_logger import htrack, hlog
-from helm.common.request import RequestResult, Sequence
+from helm.common.request import RequestResult, GeneratedOutput
 from helm.common.authentication import Authentication
 from helm.proxy.services.remote_service import RemoteService
 from helm.proxy.services.server_service import ServerService
@@ -18,28 +24,36 @@ class ExecutorError(Exception):
 
 @dataclass(frozen=True)
 class ExecutionSpec:
-    # If non-empty, URL of the proxy server we send requests to (e.g., http://localhost:1959).
+
     url: Optional[str]
+    """If non-empty, URL of the proxy server we send requests to (e.g., http://localhost:1959)."""
 
-    # Pass into the service
     auth: Authentication
+    """Authentication that will be passed into the local service, if using the local service."""
 
-    # Path where API credentials and cache is stored.
-    # This path is the same as `--base-path` when launching the proxy server (see server.py).
-    # Required when url is not set.
     local_path: Optional[str]
+    """Path where API credentials and cache is stored.
+
+    This path is the same as `--base-path` when launching the proxy server (see server.py).
+    Required when url is not set."""
 
-    # How many threads to have at once
     parallelism: int
+    """How many threads to have at once"""
 
-    # Whether to skip execution
     dry_run: bool = False
+    """Whether to skip execution"""
+
+    sqlite_cache_backend_config: Optional[SqliteCacheBackendConfig] = None
+    """If set, SQLite will be used for the cache.
+
+    This specifies the directory in which the SQLite cache will store files.
+    At most one of sqlite_cache_backend_config and mongo_cache_backend_config can be set."""
 
-    # URL to the MongoDB database.
-    # If non-empty, the MongoDB database will be used for caching instead of SQLite.
-    # Example format: mongodb://[username:password@]host1[:port1]/[dbname]
-    # For full format, see: https://www.mongodb.com/docs/manual/reference/connection-string/
-    mongo_uri: str = ""
+    mongo_cache_backend_config: Optional[MongoCacheBackendConfig] = None
+    """If set, MongoDB will be used for the cache.
+
+    This specifies the MongoDB database to be used by the MongoDB cache.
+    At most one of sqlite_cache_backend_config and mongo_cache_backend_config can be set."""
 
 
 class Executor:
@@ -51,6 +65,16 @@ class Executor:
     def __init__(self, execution_spec: ExecutionSpec):
         self.execution_spec = execution_spec
 
+        cache_backend_config: CacheBackendConfig
+        if execution_spec.sqlite_cache_backend_config and execution_spec.mongo_cache_backend_config:
+            raise ExecutorError("At most one of sqlite_cache_backend_config and mongo_cache_backend_config can be set.")
+        elif execution_spec.sqlite_cache_backend_config:
+            cache_backend_config = execution_spec.sqlite_cache_backend_config
+        elif execution_spec.mongo_cache_backend_config:
+            cache_backend_config = execution_spec.mongo_cache_backend_config
+        else:
+            cache_backend_config = BlackHoleCacheBackendConfig()
+
         self.service: Service
         if execution_spec.url:
             hlog(f"Running using remote API proxy server: {execution_spec.url}")
@@ -58,7 +82,9 @@ class Executor:
         elif execution_spec.local_path:
             hlog(f"Running in local mode with base path: {execution_spec.local_path}")
             self.service = ServerService(
-                base_path=execution_spec.local_path, root_mode=True, mongo_uri=execution_spec.mongo_uri
+                base_path=execution_spec.local_path,
+                root_mode=True,
+                cache_backend_config=cache_backend_config,
             )
         else:
             raise ValueError("Either the proxy server URL or the local path must be set")
@@ -77,7 +103,11 @@ class Executor:
         )
 
         hlog(f"Processed {len(request_states)} requests")
-        return ScenarioState(scenario_state.adapter_spec, request_states)
+        return ScenarioState(
+            adapter_spec=scenario_state.adapter_spec,
+            request_states=request_states,
+            annotator_specs=scenario_state.annotator_specs,
+        )
 
     def process(self, state: RequestState) -> RequestState:
         try:
@@ -87,7 +117,7 @@ class Executor:
         if not result.success:
             if result.error_flags and not result.error_flags.is_fatal:
                 hlog(f"WARNING: Non-fatal error treated as empty completion: {result.error}")
-                result.completions = [Sequence(text="", logprob=0, tokens=[])]
+                result.completions = [GeneratedOutput(text="", logprob=0, tokens=[])]
             else:
                 raise ExecutorError(f"{str(result.error)} Request: {state.request}")
         return replace(state, result=result)

helm/benchmark/huggingface_registration.py

@@ -4,10 +4,16 @@ from typing import Optional
 from helm.benchmark.model_deployment_registry import (
     ClientSpec,
     ModelDeployment,
-    WindowServiceSpec,
     register_model_deployment,
 )
+from helm.benchmark.model_metadata_registry import (
+    get_model_metadata,
+    get_unknown_model_metadata,
+    register_model_metadata,
+)
 from helm.benchmark.tokenizer_config_registry import TokenizerConfig, TokenizerSpec, register_tokenizer_config
+from helm.common.hierarchical_logger import hlog
+from helm.tokenizers.huggingface_tokenizer import HuggingFaceTokenizer
 
 
 def register_huggingface_model(
@@ -17,26 +23,50 @@ def register_huggingface_model(
     if revision:
         object_spec_args["revision"] = revision
 
+    # Auto-infer model properties from the tokenizer.
+    with HuggingFaceTokenizer.create_tokenizer(**object_spec_args) as tokenizer:
+        max_sequence_length = tokenizer.model_max_length
+        end_of_text_token = tokenizer.eos_token or ""
+        prefix_token = tokenizer.bos_token or ""
+    # If the tokenizer config has a model_max_length of 1000000000000000019884624838656
+    # it means that model creator did not specify model_max_length.
+    if max_sequence_length > 1_000_000:
+        raise ValueError(
+            f"Could not infer the model_max_length of Hugging Face model {pretrained_model_name_or_path}, so "
+            f"--enable-huggingface-models and --enable-local-huggingface-models cannot be used for this model. "
+            f"Please configure the model using prod_env/model_deployments.yaml instead."
+        )
+
     model_deployment = ModelDeployment(
         name=helm_model_name,
         client_spec=ClientSpec(
-            class_name="helm.proxy.clients.huggingface_client.HuggingFaceClient",
+            class_name="helm.clients.huggingface_client.HuggingFaceClient",
             args=object_spec_args,
         ),
         model_name=helm_model_name,
         tokenizer_name=helm_model_name,
-        window_service_spec=WindowServiceSpec(
-            class_name="helm.benchmark.window_services.huggingface_window_service.HuggingFaceWindowService",
-            args=object_spec_args,
-        ),
+        max_sequence_length=max_sequence_length,
     )
+
+    # We check if the model is already registered because we don't want to
+    # overwrite the model metadata if it's already registered.
+    # If it's not registered, we register it, as otherwise an error would be thrown
+    # when we try to register the model deployment.
+    try:
+        _ = get_model_metadata(model_name=helm_model_name)
+    except ValueError:
+        register_model_metadata(get_unknown_model_metadata(helm_model_name))
+        hlog(f"Registered default metadata for model {helm_model_name}")
+
     register_model_deployment(model_deployment)
     tokenizer_config = TokenizerConfig(
         name=helm_model_name,
         tokenizer_spec=TokenizerSpec(
-            class_name="helm.proxy.tokenizers.huggingface_tokenizer.HuggingFaceTokenizer",
+            class_name="helm.tokenizers.huggingface_tokenizer.HuggingFaceTokenizer",
             args=object_spec_args,
         ),
+        end_of_text_token=end_of_text_token,
+        prefix_token=prefix_token,
     )
     register_tokenizer_config(tokenizer_config)