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

helm/benchmark/scenarios/melt_ir_scenario.py

@@ -0,0 +1,171 @@
+from typing import List, Optional
+
+from datasets import load_dataset, Dataset
+from helm.common.hierarchical_logger import hlog
+from helm.benchmark.scenarios.scenario import (
+    Scenario,
+    Instance,
+    Reference,
+    TRAIN_SPLIT,
+    VALID_SPLIT,
+    CORRECT_TAG,
+    Input,
+    Output,
+    make_rank_tag,
+)
+
+
+class MELTInformationRetrievalScenario(Scenario):
+    name = "melt_information_retrieval"
+    description = "Scenario for information retrieval tasks."
+    tags = ["information_retrieval"]
+
+    """ Dictionary mapping task track tuples to the number of queries. """
+    NUM_TRAIN_QUERIES = 1000
+
+    """ Upper and lower bounds on top-k.
+
+    The top-k number represents the number of passages we will consider per
+    query. Max top-k for the train and validation files are set to the number
+    of passages included in the corresponding top-k files.
+    """
+    MIN_TOPK: int = 11
+    MAX_TRAIN_TOPK: int = 20
+    MAX_VALID_TOPK: int = 1000
+
+    def __init__(
+        self, dataset_name: str, revision: str, subset: Optional[str] = None, valid_topk: Optional[int] = None
+    ):
+        """The constructor for the MSMARCOScenario.
+
+        Args:
+            dataset_name: The name of the dataset.
+            revision: The revision of the dataset to use.
+            subset: The subset of the dataset to use. Defaults to "".
+            valid_topk: If set, specifies the number of top documents for which the
+                validation instances will be created. Must be in the range
+                [self.MIN_TOPK, self.MAX_VALID_TOPK].
+        """
+        super().__init__()
+
+        # Input validation
+        self.dataset_name = dataset_name
+        self.revision = revision
+        self.subset = subset
+        self.valid_topk: Optional[int] = valid_topk
+        if self.valid_topk is not None:
+            assert valid_topk and self.MIN_TOPK <= valid_topk <= self.MAX_VALID_TOPK
+
+    def get_train_instances(self) -> List[Instance]:
+        """Get training instances.
+        References for each instance are selected as follows:
+            1. We select 1 correct reference, where the documents included
+            corresponds to the best document for the given train query.
+            2. We create 1 wrong reference, where the document included
+            corresponds to a non-gold document for the given train query.
+        """
+        dataset = load_dataset(
+            self.dataset_name,
+            self.subset,
+            revision=self.revision,
+            trust_remote_code=True,
+        )
+        instances = []
+        for i, sample in enumerate(dataset["train"]):
+
+            if i >= self.NUM_TRAIN_QUERIES:
+                break
+
+            references = [
+                Reference(Output(text=sample["positive"]), tags=[CORRECT_TAG]),
+                Reference(Output(text=sample["negative"]), tags=[]),
+            ]
+
+            instances.append(Instance(Input(text=sample["query"]), references=references, split=TRAIN_SPLIT))
+        return instances
+
+    def get_valid_instances(self) -> List[Instance]:
+        """Get validation instances.
+        By default, we create a reference for each Document ID for which there
+        is a judgment with respect to the provided Query ID.
+
+        If self.valid_topk is not None, we ensure that a reference is created
+        for all the documents that appear in top self.valid_topk documents for
+        the given validation query.
+        """
+        dataset = load_dataset(
+            self.dataset_name,
+            f"runs-{self.subset}",
+            revision=self.revision,
+            trust_remote_code=True,
+        )
+        instances = []
+        for sample in dataset["bm25"]:
+            references = []
+
+            for k, passage_dict in enumerate(Dataset.from_dict(sample["passages"])):
+                if self.valid_topk is None or k >= self.valid_topk:
+                    break
+                tags = []
+                tags.append(f"docid={passage_dict['id']}")
+                if k == 0:
+                    tags.append(CORRECT_TAG)
+                tags.append(make_rank_tag(rank=k + 1))  # Top-k rank
+                references.append(Reference(Output(text=passage_dict["passage"]), tags=tags))
+
+            instances.append(Instance(Input(text=sample["query"]), references=references, split=VALID_SPLIT))
+
+        return instances
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        """Get instances for this scenario.
+
+        Refer to the documentation of the following methods for details on how
+        the instances are created:
+            * self.get_train_instances
+            * self.get_valid_instances
+        """
+
+        hlog("Preparing training instances.")
+        train_instances = self.get_train_instances()
+
+        hlog("Preparing validation instances.")
+        valid_instances = self.get_valid_instances()
+
+        return train_instances + valid_instances
+
+
+class MELTInformationRetrievalMMARCOScenario(MELTInformationRetrievalScenario):
+    """
+    Scenario for the MMARCO dataset.
+    """
+
+    name = "melt_information_retrieval_mmarco"
+    description = "MMARCO dataset for information retrieval in Vietnamese."
+    tags = ["information_retrieval"]
+
+    def __init__(self, **kwargs):
+        super().__init__(
+            dataset_name="unicamp-dl/mmarco",
+            revision="6d039c4638c0ba3e46a9cb7b498b145e7edc6230",
+            subset="vietnamese",
+            **kwargs,
+        )
+
+
+class MELTInformationRetrievalMRobustScenario(MELTInformationRetrievalScenario):
+    """
+    Scenario for the MRobust dataset.
+    """
+
+    name = "melt_information_retrieval_mrobust"
+    description = "MRobust dataset for information retrieval in Vietnamese."
+    tags = ["information_retrieval"]
+
+    def __init__(self, **kwargs):
+        super().__init__(
+            dataset_name="unicamp-dl/mrobust",
+            revision="fda452a7fbfd9550db2f78d9d98e6b3ec16734df",
+            subset="vietnamese",
+            **kwargs,
+        )

helm/benchmark/scenarios/melt_knowledge_scenario.py

@@ -0,0 +1,246 @@
+from abc import abstractmethod
+from typing import Dict, List, Tuple, Optional
+
+import random
+from datasets import load_dataset
+from helm.benchmark.scenarios.scenario import (
+    Scenario,
+    Instance,
+    Reference,
+    TRAIN_SPLIT,
+    TEST_SPLIT,
+    VALID_SPLIT,
+    CORRECT_TAG,
+    PassageQuestionInput,
+    Input,
+    Output,
+)
+
+
+class MELTClosedBookQAScenario(Scenario):
+    name = "melt_closed_book_qa"
+    description = "Closed Book Question Answering scenario."
+    tags = ["question_answering"]
+
+    def __init__(
+        self,
+        dataset_name: str,
+        revision: str,
+        subset: Optional[str] = None,
+        splits: Optional[Dict[str, str]] = None,
+    ):
+        """
+        Initializes the question answering scenario.
+
+        Args:
+            dataset_name: The name of the dataset.
+            revision: The revision of the dataset to use.
+            subset: The subset of the dataset to use. Defaults to "".
+            splits: The splits to use for the dataset. Defaults to None.
+        """
+        super().__init__()
+        self.dataset_name = dataset_name
+        self.subset = subset
+        self.revision = revision
+        self.splits = splits
+
+    def get_instances_for_splits(self, splits: Dict[str, str]) -> List[Instance]:
+        """
+        Helper for generating instances for a split.
+        Args:
+            splits (dict): Which splits to partition the data into.
+        Returns:
+            List[Instance]: Instances from the file for the specified split.
+        """
+        instances: List[Instance] = []
+        dataset = load_dataset(
+            self.dataset_name,
+            self.subset,
+            revision=self.revision,
+            trust_remote_code=True,
+        )
+        for dataset_split_name, helm_split_name in splits.items():
+
+            for sample in dataset[dataset_split_name]:
+                instance = Instance(
+                    input=Input(text=sample["question"]),
+                    references=[Reference(Output(text=sample["answer"]), tags=[CORRECT_TAG])],
+                    split=helm_split_name,
+                )
+                instances.append(instance)
+
+        return instances
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        if self.splits is None:
+            splits = {"train": TRAIN_SPLIT, "validation": VALID_SPLIT, "test": TEST_SPLIT}
+        else:
+            splits = {}
+            if "train" in self.splits:
+                splits[self.splits[TRAIN_SPLIT]] = TRAIN_SPLIT
+            if "validation" in self.splits:
+                splits[self.splits[VALID_SPLIT]] = VALID_SPLIT
+            if "test" in self.splits:
+                splits[self.splits[TEST_SPLIT]] = TEST_SPLIT
+
+        instances: List[Instance] = self.get_instances_for_splits(splits=splits)
+        return instances
+
+
+class MELTKnowledgeZaloScenario(MELTClosedBookQAScenario):
+    """
+    Scenario for the Zalo dataset.
+    """
+
+    name = "melt_knowledge_zalo"
+    description = "Zalo dataset for closed-book question answering."
+    tags = ["question_answering", "knowledge"]
+
+    def __init__(self):
+        super().__init__(
+            dataset_name="ura-hcmut/zalo_e2eqa",
+            revision="63494521f4de949bfa57a5f0b79bc3ee47e635ad",
+            splits={
+                TRAIN_SPLIT: "train",
+                TEST_SPLIT: "test",
+            },
+        )
+
+
+class MELTMultipleChoiceQAScenario(Scenario):
+    name = "melt_multiple_choice_qa"
+    description = "Multiple Choice Question Answering scenario."
+    tags = ["question_answering"]
+
+    def __init__(
+        self,
+        dataset_name: str,
+        revision: str,
+        subset: Optional[str] = None,
+        splits: Optional[Dict[str, str]] = None,
+    ):
+        """
+        Initializes the question answering scenario.
+
+        Args:
+            dataset_name: The name of the dataset.
+            revision: The revision of the dataset to use.
+            subset: The subset of the dataset to use. Defaults to "".
+            splits: The splits to use for the dataset. Defaults to None.
+        """
+        super().__init__()
+        self.dataset_name = dataset_name
+        self.subset = subset
+        self.revision = revision
+        self.splits = splits
+
+    @abstractmethod
+    def process_example(self, sample: dict) -> Tuple[Input, List[Reference]]:
+        """
+        Given an sample from the dataset, create the input text and
+        list of answers for the instance.
+        """
+        pass
+
+    def get_instances_for_splits(self, splits: Dict[str, str]) -> List[Instance]:
+        """
+        Helper for generating instances for a split.
+        Args:
+            splits (dict): Which splits to partition the data into.
+        Returns:
+            List[Instance]: Instances from the file for the specified split.
+        """
+        instances: List[Instance] = []
+        dataset = load_dataset(
+            self.dataset_name,
+            self.subset,
+            revision=self.revision,
+            trust_remote_code=True,
+        )
+        for dataset_split_name, helm_split_name in splits.items():
+            for sample in dataset[dataset_split_name]:
+                inputs, references = self.process_example(sample)
+                instance = Instance(
+                    input=inputs,
+                    references=references,
+                    split=helm_split_name,
+                )
+                instances.append(instance)
+
+        return instances
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        if self.splits is None:
+            splits = {"train": TRAIN_SPLIT, "validation": VALID_SPLIT, "test": TEST_SPLIT}
+        else:
+            splits = {}
+            if "train" in self.splits:
+                splits[self.splits[TRAIN_SPLIT]] = TRAIN_SPLIT
+            if "validation" in self.splits:
+                splits[self.splits[VALID_SPLIT]] = VALID_SPLIT
+            if "test" in self.splits:
+                splits[self.splits[TEST_SPLIT]] = TEST_SPLIT
+
+        instances: List[Instance] = self.get_instances_for_splits(splits=splits)
+        return instances
+
+
+class MELTKnowledgeViMMRCScenario(MELTMultipleChoiceQAScenario):
+    """
+    Scenario for the ViMMRC dataset.
+    """
+
+    name = "melt_knowledge_vimmrc"
+    description = "ViMMRC dataset for multiple choice question answering."
+    tags = ["question_answering", "knowledge"]
+
+    def __init__(self, randomize_order: bool = False):
+        super().__init__(
+            dataset_name="ura-hcmut/ViMMRC",
+            revision="fe68800e37aaa84d80b1d93466b36c3fa60d8bcb",
+            splits={
+                TRAIN_SPLIT: "train",
+                VALID_SPLIT: "validation",
+                TEST_SPLIT: "test",
+            },
+        )
+        self.randomize_order = randomize_order
+        self.correct_answer_mapping = {
+            "A": 0,
+            "B": 1,
+            "C": 2,
+            "D": 3,
+            "E": 4,
+            "F": 5,
+            "G": 6,
+            "H": 7,
+        }
+        random.seed(42)
+
+    def process_example(self, sample: dict) -> Tuple[Input, List[Reference]]:
+        """
+        Given an sample from the dataset, create the input text and
+        list of answers for the instance.
+        """
+        inputs = PassageQuestionInput(
+            passage=sample["article"],
+            passage_prefix="Ngữ cảnh: ",
+            question=sample["question"],
+            question_prefix="Câu hỏi: ",
+            separator="\n\n",
+        )
+
+        correct_idx = self.correct_answer_mapping[sample["answer"]]
+        references = []
+        for idx, answer in enumerate(eval(sample["options"])):
+            if idx == correct_idx:
+                tags = [CORRECT_TAG]
+            else:
+                tags = []
+
+            references.append(Reference(Output(text=answer), tags=tags))
+
+        if self.randomize_order:
+            random.shuffle(references)
+
+        return inputs, references

helm/benchmark/scenarios/melt_lm_scenarios.py

@@ -0,0 +1,252 @@
+from typing import Any, Dict, List, Optional
+
+import random
+from datasets import load_dataset
+from helm.benchmark.scenarios.scenario import (
+    Scenario,
+    Instance,
+    Reference,
+    TRAIN_SPLIT,
+    TEST_SPLIT,
+    VALID_SPLIT,
+    CORRECT_TAG,
+    Input,
+    Output,
+)
+
+
+class MELTLMMaskFillingScenario(Scenario):
+    """
+    Scenario for the MELT Masked Language Modeling dataset.
+    """
+
+    name = "melt_lm_mask_filling"
+    description = "Masked Language Modeling scenario."
+    tags = ["language_modeling", "mask_filling"]
+
+    def __init__(
+        self,
+        dataset_name: str,
+        revision: str,
+        masked_ratio: float = 0.1,
+        text_key: str = "text",
+        subset: Optional[str] = None,
+        splits: Optional[Dict[str, str]] = None,
+    ):
+        """Initializes the question answering scenario.
+
+        Args:
+            dataset_name: The name of the dataset.
+            revision: The revision of the dataset to use.
+            masked_ratio: The ratio of tokens to mask in the input text. Defaults to 0.1.
+            text_key: The key to use for the text in the dataset. Defaults to "text".
+            subset: The subset of the dataset to use. Defaults to "".
+            splits: The splits to use for the dataset. Defaults to None.
+        """
+        super().__init__()
+        self.dataset_name = dataset_name
+        self.subset = subset
+        self.masked_ratio = masked_ratio
+        self.text_key = text_key
+        self.revision = revision
+        self.splits = splits
+
+    def _mask_text(self, text: str) -> str:
+        """
+        Mask a portion of the input text.
+        Args:
+            text (str): The input text to mask.
+        Returns:
+            str: The masked text.
+        """
+        tokens = text.split(" ")
+        num_tokens_to_mask = int(len(tokens) * self.masked_ratio)
+        indices_to_mask = random.sample(range(len(tokens)), num_tokens_to_mask)
+        for index in indices_to_mask:
+            tokens[index] = "[MASK]"
+        return " ".join(tokens)
+
+    def get_instances_for_splits(self, splits: Dict[str, str]) -> List[Instance]:
+        """
+        Helper for generating instances for a split.
+        Args:
+            splits (dict): Which splits to partition the data into.
+        Returns:
+            List[Instance]: Instances from the file for the specified split.
+        """
+        instances: List[Instance] = []
+        dataset: Any = load_dataset(
+            self.dataset_name,
+            self.subset,
+            revision=self.revision,
+            trust_remote_code=True,
+        )
+
+        for dataset_split_name, helm_split_name in splits.items():
+            for sample in dataset[dataset_split_name]:
+                target_sentence = sample[self.text_key]
+                source_sentence = self._mask_text(target_sentence)
+                instances.append(
+                    Instance(
+                        input=Input(text=source_sentence),
+                        references=[Reference(Output(text=target_sentence), tags=[CORRECT_TAG])],
+                        split=helm_split_name,
+                    )
+                )
+        return instances
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        if self.splits is None:
+            splits = {"train": TRAIN_SPLIT, "validation": VALID_SPLIT, "test": TEST_SPLIT}
+        else:
+            splits = {}
+            if "train" in self.splits:
+                splits[self.splits[TRAIN_SPLIT]] = TRAIN_SPLIT
+            if "validation" in self.splits:
+                splits[self.splits[VALID_SPLIT]] = VALID_SPLIT
+            if "test" in self.splits:
+                splits[self.splits[TEST_SPLIT]] = TEST_SPLIT
+
+        random.seed(42)
+        instances: List[Instance] = self.get_instances_for_splits(splits=splits)
+        return instances
+
+
+class MELTLMMaskFillingMLQAScenario(MELTLMMaskFillingScenario):
+    """
+    Scenario for the MLQA dataset.
+    This dataset is a multilingual question answering dataset.
+    It contains questions in multiple languages and their corresponding
+    answers in the same language. In this scenario, we are using the
+    context of questions in the Vietnamese subset of the MLQA dataset.
+    """
+
+    name = "melt_lm_mask_filling_mlqa"
+    description = "MLQA dataset for masked language modeling."
+    tags = ["language_modeling", "mask_filling"]
+
+    def __init__(self):
+        super().__init__(
+            dataset_name="facebook/mlqa",
+            revision="397ed406c1a7902140303e7faf60fff35b58d285",
+            subset="mlqa.vi.vi",
+            text_key="context",
+            splits={
+                VALID_SPLIT: "validation",
+                TEST_SPLIT: "test",
+            },
+        )
+
+
+class MELTLMSpellingCorrectionScenario(Scenario):
+    """
+    Scenario for the MELT spelling correction dataset.
+    """
+
+    name = "melt_lm_spelling_correction"
+    description = "Spelling Correction scenario."
+    tags = ["language_modeling", "spelling_correction"]
+
+    def __init__(
+        self,
+        dataset_name: str,
+        revision: str,
+        source_key: str = "text",
+        target_key: str = "corrected_text",
+        subset: Optional[str] = None,
+        splits: Optional[Dict[str, str]] = None,
+    ):
+        """Initializes the question answering scenario.
+
+        Args:
+            dataset_name: The name of the dataset.
+            revision: The revision of the dataset to use.
+            source_key: The key to use for the source text in the dataset. Defaults to "text".
+            target_key: The key to use for the target text in the dataset. Defaults to "corrected_text".
+            subset: The subset of the dataset to use. Defaults to "".
+            splits: The splits to use for the dataset. Defaults to None.
+        """
+        super().__init__()
+        self.dataset_name = dataset_name
+        self.subset = subset
+        self.source_key = source_key
+        self.target_key = target_key
+        self.revision = revision
+        self.splits = splits
+
+    def get_instances_for_splits(self, splits: Dict[str, str]) -> List[Instance]:
+        """
+        Helper for generating instances for a split.
+        Args:
+            splits (dict): Which splits to partition the data into.
+        Returns:
+            List[Instance]: Instances from the file for the specified split.
+        """
+        instances: List[Instance] = []
+        dataset: Any = load_dataset(
+            self.dataset_name,
+            self.subset,
+            revision=self.revision,
+            trust_remote_code=True,
+        )
+        if len(splits) == 1:
+            all_keys = list(splits.keys())
+            dataset = dataset[all_keys[0]].train_test_split(test_size=0.33, seed=42)
+            splits = {
+                "train": TRAIN_SPLIT,
+                "test": TEST_SPLIT,
+            }
+
+        for dataset_split_name, helm_split_name in splits.items():
+            for sample in dataset[dataset_split_name]:
+                source_sentence = sample[self.source_key]
+                target_sentence = sample[self.target_key]
+                instances.append(
+                    Instance(
+                        input=Input(text=source_sentence),
+                        references=[Reference(Output(text=target_sentence), tags=[CORRECT_TAG])],
+                        split=helm_split_name,
+                    )
+                )
+        return instances
+
+    def get_instances(self, output_path: str) -> List[Instance]:
+        if self.splits is None:
+            splits = {"train": TRAIN_SPLIT, "validation": VALID_SPLIT, "test": TEST_SPLIT}
+        else:
+            splits = {}
+            if "train" in self.splits:
+                splits[self.splits[TRAIN_SPLIT]] = TRAIN_SPLIT
+            if "validation" in self.splits:
+                splits[self.splits[VALID_SPLIT]] = VALID_SPLIT
+            if "test" in self.splits:
+                splits[self.splits[TEST_SPLIT]] = TEST_SPLIT
+
+        random.seed(42)
+        instances: List[Instance] = self.get_instances_for_splits(splits=splits)
+        return instances
+
+
+class MELTLMSpellingCorrectionVSECScenario(MELTLMSpellingCorrectionScenario):
+    """
+    Scenario for the VSEC dataset.
+    The VSEC dataset is a Vietnamese spelling correction dataset.
+    It contains 9,341 pairs of sentences where the first sentence is a misspelled
+    version of the second sentence, which is the correct version.
+    The mistakes are common spelling errors made by Vietnamese speakers and typists.
+    """
+
+    name = "melt_lm_spelling_correction_vsec"
+    description = "VSEC dataset for spelling correction."
+    tags = ["language_modeling", "spelling_correction"]
+
+    def __init__(self):
+        super().__init__(
+            dataset_name="ura-hcmut/VSEC",
+            revision="a6732e131605b5ec24ecc1745c6061c5ae86814e",
+            source_key="text",
+            target_key="correct",
+            splits={
+                TEST_SPLIT: "test",
+            },
+        )