eval-framework 0.2.0__py3-none-any.whl

eval_framework/shared/types.py

@@ -0,0 +1,227 @@
+import re
+from collections.abc import Callable, Sequence
+from typing import Annotated, NamedTuple, Self, TypeVar, cast
+
+from pydantic import BaseModel, ConfigDict
+
+from eval_framework.metrics.llm.graders.language import detect_language_of
+from eval_framework.utils.helpers import count_bytes
+from template_formatting.formatter import ConcatFormatter, Message, Role
+
+
+class ConcatCompression(NamedTuple):
+    """Helper class for storing compression info for the concat formatter.
+
+    The concat formatter is used to avoid bias towards special tokens.
+    """
+
+    num_bytes: int
+    num_tokens: int
+
+    @classmethod
+    def calculate(
+        cls,
+        messages: Sequence[Message],
+        count_tokens: Callable[[str], int],
+        choices: list[str] | None = None,
+        completion: str | None = None,
+    ) -> Self | None:
+        """Calculate the compression info for the given messages and token counting function."""
+        if (choices is None) == (completion is None):
+            raise ValueError("Either possible_completions or completion must be provided, but not both.")
+        concat_str = ConcatFormatter().format(messages, output_mode="string")
+
+        if choices is not None:
+            if any(c is None for c in choices):
+                return None
+            num_bytes = count_bytes(concat_str) + sum(count_bytes(c) for c in choices)
+            num_tokens = count_tokens(concat_str) + sum(count_tokens(c) for c in choices)
+        else:
+            if completion is None:
+                return None
+            concat_str = f"{concat_str}{completion}"
+            num_bytes = count_bytes(concat_str)
+            num_tokens = count_tokens(concat_str)
+
+        res = cls(num_bytes=num_bytes, num_tokens=num_tokens)
+        if res.num_bytes > 0 and res.num_tokens > 0:
+            return res
+        else:
+            return None
+
+
+class BaseMetricContext(BaseModel):
+    """Base class for metric context"""
+
+    model_config = ConfigDict(extra="allow")
+
+
+class LanguageMetricContext(BaseMetricContext):
+    language: str
+
+
+class UntemplatedPrompt(BaseMetricContext):
+    untemplated_prompt: str
+
+
+class Error(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    error_class: str
+    message: str
+    traceback: str
+
+
+class PromptTooLongException(Exception):
+    pass
+
+
+class BaseCompletion(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    prompt: Annotated[str, "prompt as passed to the llm"]
+    prompt_sequence_positions: Annotated[
+        int | None,
+        "number of sequence positions that the prompt occupies in the llm architecture (e.g. token count) "
+        "or None if the info is not available",
+    ]
+    completion: Annotated[str, "completion as generated by the llm"]
+    concat_compression: Annotated[ConcatCompression | None, "Compression info for the concat formatter."] = None
+
+
+class RawCompletion(BaseCompletion):
+    completion_sequence_positions: Annotated[
+        int | None,
+        "number of sequence positions that the completion occupies in the llm architecture "
+        "(e.g. token count) or None if the info is not available",
+    ]
+    raw_completion_error: Error | None = None
+
+
+class Completion(BaseCompletion):
+    id: int
+    subject: str
+    ground_truth: str | None | list[str]
+    messages: list[Message] | None  # needed for LLM as a judge
+    raw_completion: Annotated[str, "raw completion as generated by the llm"]
+    raw_completion_sequence_positions: Annotated[
+        int | None,
+        "number of sequence positions that the completion occupies in the llm architecture or None "
+        "if the info is not available",
+    ]
+    context: list[BaseMetricContext] | BaseMetricContext | None = None
+    error: Error | None = None
+
+    @property
+    def ground_truth_list(self) -> list[str] | list[None]:
+        if isinstance(self.ground_truth, list):
+            return self.ground_truth
+
+        return [self.ground_truth]  # type: ignore[return-value]
+
+    # Use just the raw messages for instructions to LLM judges, not the original prompt with its special formatting.
+    # (see https://x.com/karpathy/status/1823418177197646104 for a motivation).
+    @property
+    def system_user_instruction(self) -> str:
+        assert self.messages is not None
+        return "\n\n".join([m.content for m in self.messages if m.role in (Role.SYSTEM, Role.USER)])
+
+    @property
+    def user_instruction(self) -> str:
+        assert self.messages is not None
+        return "\n\n".join([m.content for m in self.messages if m.role == Role.USER])
+
+    @property
+    def first_user_instruction(self) -> str:
+        assert self.messages is not None
+        user_messages = [m.content for m in self.messages if m.role == Role.USER]
+        return user_messages[0] if user_messages else ""
+
+    @property
+    def all_but_first_user_instruction(self) -> str:
+        assert self.messages is not None
+        user_messages = [m.content for m in self.messages if m.role == Role.USER]
+        return "\n\n".join(user_messages[1:]) if len(user_messages) > 1 else ""
+
+    @property
+    def last_user_instruction(self) -> str:
+        assert self.messages is not None
+        user_messages = [m.content for m in self.messages if m.role == Role.USER]
+        return user_messages[-1] if user_messages else ""
+
+    @property
+    def sanitized_completion(self) -> str:
+        # Make sure the completion doesn't contain any obvious special chars either by "breaking" any <|xyz|> pattern.
+        return re.sub(r"<\|(\S+)\|>", r"<| \1 |>", self.completion)
+
+    def get_completion_language(self) -> str:
+        detected_language = ""
+        if self.context and isinstance(self.context, LanguageMetricContext):
+            detected_language = self.context.language
+        else:
+            detected_language_object = detect_language_of(self.completion)
+            detected_language = detected_language_object.iso_code_639_1.name.lower() if detected_language_object else ""
+        return detected_language
+
+    def get_raw_completion_language(self) -> str:
+        detected_language = ""
+        if self.context and isinstance(self.context, LanguageMetricContext):
+            detected_language = self.context.language
+        else:
+            detected_language_object = detect_language_of(self.raw_completion)
+            detected_language = detected_language_object.iso_code_639_1.name.lower() if detected_language_object else ""
+        return detected_language
+
+    def get_instruction_language(self) -> str:
+        detected_language = ""
+        if self.context and isinstance(self.context, LanguageMetricContext):
+            detected_language = self.context.language
+        else:
+            detected_language_object = detect_language_of(self.user_instruction)
+            detected_language = detected_language_object.iso_code_639_1.name.lower() if detected_language_object else ""
+        return detected_language
+
+
+class BaseLoglikelihood(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    prompt: str
+    prompt_sequence_positions: int | None
+    loglikelihoods: dict[str, float]
+    loglikelihoods_sequence_positions: dict[str, int]  # Is empty if the model does not provide sequence positions
+    concat_compression: Annotated[ConcatCompression | None, "Compression info for the concat formatter"] = None
+
+
+class RawLoglikelihood(BaseLoglikelihood):
+    raw_loglikelihood_error: Error | None = None
+
+
+class Loglikelihood(BaseLoglikelihood):
+    id: int
+    subject: str
+    ground_truth: str | list[str]
+    error: Error | None = None
+
+    @property
+    def ground_truth_list(self) -> list[str] | list[None]:
+        if isinstance(self.ground_truth, list):
+            return self.ground_truth
+        return [self.ground_truth]  # type: ignore[return-value]
+
+
+MetricContext = TypeVar("MetricContext", bound=BaseMetricContext)
+
+
+def extract_context_metric[MetricContext: BaseMetricContext](
+    response: Completion, metric_context_class: type[MetricContext]
+) -> MetricContext:
+    assert response.context is not None, "Expected context to be provided in the response"
+    if not isinstance(response.context, list):
+        assert isinstance(response.context, metric_context_class) or isinstance(response.context, BaseMetricContext), (
+            f"Expected context to be of type {metric_context_class.__name__}, got {type(response.context).__name__}"
+        )
+        return cast(MetricContext, response.context)
+    else:
+        assert len(response.context) > 0, "Expected context to be provided in the response"
+        context = [
+            metric_context for metric_context in response.context if isinstance(metric_context, metric_context_class)
+        ][0]
+        assert context is not None, f"Expected {metric_context_class.__name__} to be provided in the response context"
+        return cast(MetricContext, context)

eval_framework/tasks/__init__.py

@@ -0,0 +1,6 @@
+# Register all tasks on import
+from .task_names import register_all_tasks
+
+register_all_tasks()
+
+del register_all_tasks

eval_framework/tasks/base.py

@@ -0,0 +1,314 @@
+import logging
+import os
+import random
+from abc import ABC, abstractmethod
+from collections.abc import Iterable
+from enum import Enum
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, Self, TypeVar
+
+import iso639
+from datasets import DatasetDict, DownloadConfig, load_dataset
+from huggingface_hub import HfApi
+from huggingface_hub.errors import RevisionNotFoundError
+from pydantic import BaseModel, ConfigDict
+
+from eval_framework.shared.types import BaseMetricContext
+from template_formatting.formatter import Message, Role
+
+if TYPE_CHECKING:
+    from eval_framework.metrics.base import BaseMetric
+
+RANDOM_SEED = 42
+NO_SUBJECT = "no_subject"
+
+
+class ResponseType(Enum):
+    COMPLETION = "completion"
+    LOGLIKELIHOODS = "loglikelihoods"
+
+
+class Language(Enum):
+    ENG = "English"
+    DEU = "German"
+    FRA = "French"
+    ITA = "Italian"
+    SPA = "Spanish"
+    POR = "Portuguese"
+    NLD = "Dutch"
+    FIN = "Finnish"
+    SWE = "Swedish"
+    ARB = "Arabic"
+    POL = "Polish"
+    RUS = "Russian"
+    UKR = "Ukrainian"
+    HRV = "Croatian"
+    SRP = "Serbian"
+
+    @classmethod
+    def add_members(cls, new_members: dict[str, Any]) -> type["Language"]:
+        members = {member.name: member.value for member in cls}
+        for name, value in new_members.items():
+            if name not in members:
+                members[name] = value
+        return Enum(cls.__name__, members)  # type: ignore[return-value]
+
+
+languages: dict[str, str] = {}
+for language in iso639.ALL_LANGUAGES:
+    enum_name = language.part3.upper()
+    languages[enum_name] = language.name
+
+Language: type[Enum] = Language.add_members(languages)  # type: ignore[no-redef]
+
+
+class Sample(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+    id: int
+    subject: str
+    messages: list[Message]
+    ground_truth: str | list[str] | None
+    possible_completions: list[str] | None
+    context: BaseMetricContext | list[BaseMetricContext] | None = None
+
+
+SubjectType = TypeVar("SubjectType")
+
+logger = logging.getLogger(__name__)
+
+
+class BaseTask[SubjectType](ABC):
+    NAME: str
+    DATASET_PATH: str
+    SAMPLE_SPLIT: str
+    FEWSHOT_SPLIT: str
+    RESPONSE_TYPE: ResponseType
+    METRICS: list[type["BaseMetric"]]
+    SUBJECTS: list[SubjectType]
+    HF_REVISION: str | None = None  # tag name, or branch name, or commit hash to ensure reproducibility
+
+    # Words in _get_instruction_text() not to be perturbed. List of words is case insensitive. No special characters
+    # or whitespace should be included.
+    PERTURBATION_UNMODIFIABLE_WORDS: list[str] | None
+
+    # The language (or languages) tested by the benchmark. Accepts a single string, a dictionary specifying
+    # language by subtopic, or `None` (for tasks not specific to a single language).
+    LANGUAGE: Language | dict[str, Language] | dict[str, tuple[Language, Language]] | None
+
+    def __init__(self, num_fewshot: int = 0) -> None:
+        self.num_fewshot = num_fewshot
+        self.stop_sequences: list[str] | None = None
+        self.max_tokens: int | None = None
+
+    @classmethod
+    def with_overwrite(
+        cls, num_fewshot: int, *, custom_subjects: list[str] | None, custom_hf_revision: str | None
+    ) -> Self:
+        instance = cls(num_fewshot=num_fewshot)
+
+        # If custom subjects were provided during initialization, they take precedence over the class-level SUBJECTS.
+        filtered_subjects = instance._filter_task_subjects(custom_subjects=custom_subjects)
+        if filtered_subjects:
+            logger.info(f"Setting SUBJECTS to `{filtered_subjects}` for the task {instance.__class__.__name__}")
+            instance.SUBJECTS = filtered_subjects  # type: ignore[assignment]
+
+        # If a custom revision was provided during initialization, it takes precedence over the class-level HF_REVISION.
+        if custom_hf_revision:
+            logger.info(f"Setting HF revision to `{custom_hf_revision}` for the task {instance.__class__.__name__}")
+            instance.HF_REVISION = custom_hf_revision
+
+        return instance
+
+    def _filter_task_subjects(self, custom_subjects: list[str] | None) -> list[str] | list[tuple] | None:
+        """Process custom subjects passed from EvalConfig. Check and returns restricted task subjects if specified."""
+        if not custom_subjects:
+            return None
+
+        assert hasattr(self, "SUBJECTS") and len(self.SUBJECTS) > 0
+        if isinstance(self.SUBJECTS[0], tuple):
+            # subjects are specified as strings but we need tuples
+            filters = [tuple(item.strip() for item in subject.split(",")) for subject in custom_subjects]
+
+            # check if all parts of custom subjects exists (* is a wildcard)
+            num_items = len(self.SUBJECTS[0])
+            legal_values = [
+                set([s[i] for s in self.SUBJECTS if isinstance(s, tuple)] + ["*"]) for i in range(num_items)
+            ]
+
+            for tpl in filters:
+                for i, v in enumerate(tpl):
+                    assert v in legal_values[i], f"Subject part {v} not found in task {self.__class__.__name__}"
+
+            # filter task subjects. * is a supported wildcard for a specific item in a tuple, e.g. "DE_DE, *"
+            chosen_subjects = []
+            for subject in self.SUBJECTS:
+                subject_tuple = subject if isinstance(subject, tuple) else tuple(str(subject).split(","))
+                for filter in filters:
+                    if all(filter[i] == "*" or filter[i] == subject_tuple[i] for i in range(num_items)):
+                        chosen_subjects.append(subject_tuple)
+                        break
+            return chosen_subjects  # type: ignore[return-value]
+        else:
+            for cs in custom_subjects:
+                assert cs in self.SUBJECTS, f"Subject {cs} not found in task {self.__class__.__name__}"
+            return custom_subjects  # type: ignore[return-value]
+
+    def _load_hf_dataset(self, **kwargs: Any) -> Any:
+        # Check if the HF_REVISION is valid before loading the dataset
+        if self.HF_REVISION:
+            try:
+                _ = HfApi().dataset_info(repo_id=kwargs["path"], revision=self.HF_REVISION, timeout=100.0)
+            except Exception as e:
+                if isinstance(e, RevisionNotFoundError):
+                    raise e
+
+        cache_dir: str = os.environ.get("HF_DATASET_CACHE_DIR", f"{Path.home()}/.cache/huggingface/datasets")
+        download_config = DownloadConfig(cache_dir=cache_dir, max_retries=5)
+        try:
+            return load_dataset(
+                **kwargs,
+                revision=self.HF_REVISION,
+                trust_remote_code=True,
+                cache_dir=cache_dir,
+                download_config=download_config,
+            )
+        except Exception:
+            return load_dataset(
+                **kwargs,
+                revision=self.HF_REVISION,
+                trust_remote_code=True,
+                cache_dir=f"{Path.home()}/.cache/eval-framework",
+            )
+
+    def _shuffle_splits(self, hf_dataset: DatasetDict) -> dict[str, list[dict[str, Any]]]:
+        dataset = {}
+        self.rnd = random.Random(RANDOM_SEED)
+
+        for split, data in hf_dataset.items():
+            if split not in [self.SAMPLE_SPLIT, self.FEWSHOT_SPLIT]:
+                continue
+
+            data_list = list(data)
+
+            if split == self.SAMPLE_SPLIT:
+                self.rnd.shuffle(data_list)
+
+            dataset[split] = data_list
+
+        return dataset
+
+    def _load_dataset(self, subject: SubjectType) -> None:
+        name = subject if subject != NO_SUBJECT else None
+        hf_dataset = self._load_hf_dataset(path=self.DATASET_PATH, name=name)
+        self.dataset = self._shuffle_splits(hf_dataset=hf_dataset)
+
+    def post_process_generated_completion(self, completion_text: str, sample: Sample | None = None) -> str:
+        return completion_text
+
+    def _get_example_messages(self, item: dict[str, Any]) -> list[Message]:
+        fewshot_examples = self._sample_fewshot_examples(item) if self.num_fewshot > 0 else []
+
+        example_messages = []
+        for fewshot_example in fewshot_examples:
+            fewshot_example["subject"] = item["subject"]
+            example_messages.extend(self._get_instruction_messages(fewshot_example))
+            example_messages.append(
+                Message(role=Role.ASSISTANT, content=self._get_fewshot_target_text(fewshot_example))
+            )
+        return example_messages
+
+    def _get_messages(self, item: dict[str, Any]) -> list[Message]:
+        example_messages = self._get_example_messages(item)
+        instruction_message = self._get_instruction_messages(item)
+        cue_text = self._get_cue_text(item)
+        cue_message = [Message(role=Role.ASSISTANT, content=cue_text)] if cue_text else []
+        messages = example_messages + instruction_message + cue_message
+        if initial_prompt_text := self._get_initial_prompt_text(item):
+            first_message = messages[0]
+            assert first_message.role == Role.USER
+            first_message.content = f"{initial_prompt_text}\n\n{first_message.content}"
+
+        if system_prompt_text := self._get_system_prompt_text(item):
+            return [Message(role=Role.SYSTEM, content=system_prompt_text)] + messages
+        return messages
+
+    def _get_instruction_messages(self, item: dict[str, Any]) -> list[Message]:
+        return [Message(role=Role.USER, content=self._get_instruction_text(item))]
+
+    def iterate_samples(self, num_samples: int | None = None) -> Iterable[Sample]:
+        for subject in self.SUBJECTS:
+            self._load_dataset(subject)
+            assert len(self.dataset[self.SAMPLE_SPLIT]) > 0
+            done = False
+            index = 0
+            for item in self.dataset[self.SAMPLE_SPLIT]:
+                if done:
+                    break
+                item["subject"] = subject
+                for sample in self._create_samples(item, index, str(subject)):
+                    yield sample
+                    index += 1
+                    if index == num_samples:
+                        done = True
+                        break
+
+    def _create_samples(self, item: dict[str, Any], index: int, subject: str) -> list[Sample]:
+        """Creates one or more samples from a single dataset item. Default implementation returns single sample."""
+        return [
+            Sample(
+                id=index,
+                subject=str(subject),
+                messages=self._get_messages(item),
+                ground_truth=self._get_ground_truth(item),
+                possible_completions=self._get_possible_completions(item),
+                context=self._get_context(item),
+            )
+        ]
+
+    def _get_initial_prompt_text(self, item: dict[str, Any]) -> str:
+        return ""
+
+    def _get_system_prompt_text(self, item: dict[str, Any]) -> str | None:
+        return None
+
+    @abstractmethod
+    def _get_instruction_text(self, item: dict[str, Any]) -> str:
+        raise NotImplementedError
+
+    def _get_fewshot_target_text(self, item: dict[str, Any]) -> str:
+        target = self._get_ground_truth(item)
+        assert target is not None
+        assert isinstance(target, str)
+        return target
+
+    @abstractmethod
+    def _get_ground_truth(self, item: dict[str, Any]) -> str | None | list[str]:
+        raise NotImplementedError
+
+    def _get_cue_text(self, item: dict[str, Any]) -> str:
+        return ""
+
+    def _get_possible_completions(self, item: dict[str, Any]) -> list[str] | None:
+        return None
+
+    def _sample_fewshot_examples(self, item: dict[str, Any]) -> list[dict]:
+        if self.FEWSHOT_SPLIT == self.SAMPLE_SPLIT:
+            fewshot_examples = self.rnd.sample(self.dataset[self.FEWSHOT_SPLIT], self.num_fewshot + 1)
+            fewshot_examples = [example for example in fewshot_examples if example != item]
+            fewshot_examples = fewshot_examples[: self.num_fewshot]
+            return fewshot_examples
+        else:
+            return self.rnd.sample(self.dataset[self.FEWSHOT_SPLIT], self.num_fewshot)
+
+    def _get_context(self, item: dict[str, Any]) -> BaseMetricContext | list[BaseMetricContext] | None:
+        return None
+
+    def get_metadata(self) -> dict[str, str | list[str]]:
+        return {
+            "dataset_path": self.DATASET_PATH,
+            "sample_split": self.SAMPLE_SPLIT,
+            "fewshot_split": self.FEWSHOT_SPLIT,
+            "response_type": self.RESPONSE_TYPE.value,
+            "metrics": [m.NAME for m in self.METRICS],
+            "subjects": [str(s) for s in self.SUBJECTS],
+        }

eval_framework/tasks/benchmarks/arc.py

@@ -0,0 +1,46 @@
+from typing import Any
+
+from eval_framework.metrics.loglikelihood.accuracy_loglikelihood import (
+    AccuracyLoglikelihood,
+    AccuracyNormLoglikelihood,
+)
+from eval_framework.tasks.base import BaseTask, Language, ResponseType
+from eval_framework.tasks.utils import get_n_letters
+
+
+class ARC(BaseTask[str]):
+    """ARC dataset: https://huggingface.co/datasets/allenai/ai2_arc"""
+
+    NAME = "ARC"
+    DATASET_PATH = "ai2_arc"
+    SAMPLE_SPLIT = "test"
+    FEWSHOT_SPLIT = "train"
+    RESPONSE_TYPE = ResponseType.LOGLIKELIHOODS
+    METRICS = [AccuracyLoglikelihood, AccuracyNormLoglikelihood]
+    SUBJECTS = ["ARC-Easy", "ARC-Challenge"]
+    PERTURBATION_UNMODIFIABLE_WORDS = ["Question"] + get_n_letters(5)
+    LANGUAGE = Language.ENG
+
+    def __init__(self, num_fewshot: int = 0) -> None:
+        super().__init__(num_fewshot)
+
+        self.keys = get_n_letters(5)  # needs to be 5 because there is one sample with 5 answer possibilities
+        self.num_to_letter = {str(i): letter for i, letter in enumerate(self.keys, start=1)}
+
+    def _get_instruction_text(self, item: dict[str, Any]) -> str:
+        return f"Question: {item['question']}\n"
+
+    def _get_fewshot_target_text(self, item: dict[str, Any]) -> str:
+        ground_truth = self._get_ground_truth(item)
+        assert ground_truth is not None
+        return f"{self._get_cue_text(item)}{ground_truth}"
+
+    def _get_cue_text(self, item: dict[str, Any]) -> str:
+        return "Answer:"
+
+    def _get_ground_truth(self, item: dict[str, Any]) -> str | None:
+        answer_key = self.num_to_letter.get(item["answerKey"], item["answerKey"])
+        return f" {item['choices']['text'][self.keys.index(answer_key)]}"
+
+    def _get_possible_completions(self, item: dict[str, Any]) -> list[str] | None:
+        return [f" {choice}" for choice in item["choices"]["text"]]

eval_framework/tasks/benchmarks/arc_de.py

@@ -0,0 +1,46 @@
+from typing import Any
+
+from eval_framework.metrics.loglikelihood.accuracy_loglikelihood import (
+    AccuracyLoglikelihood,
+    AccuracyNormLoglikelihood,
+)
+from eval_framework.tasks.base import NO_SUBJECT, BaseTask, Language, ResponseType
+from eval_framework.tasks.utils import get_n_letters
+
+
+class ARC_DE(BaseTask[str]):
+    """ARC-DE dataset: https://huggingface.co/datasets/LeoLM/ArcChallenge_de"""
+
+    NAME = "ARC German"
+    DATASET_PATH = "LeoLM/ArcChallenge_de"
+    SAMPLE_SPLIT = "test"
+    FEWSHOT_SPLIT = "validation"
+    RESPONSE_TYPE = ResponseType.LOGLIKELIHOODS
+    METRICS = [AccuracyLoglikelihood, AccuracyNormLoglikelihood]
+    SUBJECTS = [NO_SUBJECT]
+    PERTURBATION_UNMODIFIABLE_WORDS = ["Frage"] + get_n_letters(5)
+    LANGUAGE = Language.DEU
+
+    def __init__(self, num_fewshot: int = 0) -> None:
+        super().__init__(num_fewshot)
+
+        self.keys = get_n_letters(5)  # needs to be 5 because there is one sample with 5 answer possibilities
+        self.num_to_letter = {str(i): letter for i, letter in enumerate(self.keys, start=1)}
+
+    def _get_instruction_text(self, item: dict[str, Any]) -> str:
+        return f"Frage: {item['question_de']}\n"
+
+    def _get_fewshot_target_text(self, item: dict[str, Any]) -> str:
+        ground_truth = self._get_ground_truth(item)
+        assert ground_truth is not None
+        return f"{self._get_cue_text(item)}{ground_truth}"
+
+    def _get_cue_text(self, item: dict[str, Any]) -> str:
+        return "Antwort:"
+
+    def _get_ground_truth(self, item: dict[str, Any]) -> str | None:
+        answer_key = self.num_to_letter.get(item["answerKey"], item["answerKey"])
+        return f" {item['choices_de']['text'][self.keys.index(answer_key)]}"
+
+    def _get_possible_completions(self, item: dict[str, Any]) -> list[str] | None:
+        return [f" {choice}" for choice in item["choices_de"]["text"]]

eval_framework/tasks/benchmarks/arc_fi.py

@@ -0,0 +1,46 @@
+from typing import Any
+
+from eval_framework.metrics.loglikelihood.accuracy_loglikelihood import (
+    AccuracyLoglikelihood,
+    AccuracyNormLoglikelihood,
+)
+from eval_framework.tasks.base import BaseTask, Language, ResponseType
+from eval_framework.tasks.utils import get_n_letters
+
+
+class ARC_FI(BaseTask[str]):
+    """ARC-FI dataset: https://huggingface.co/datasets/LumiOpen/arc_challenge_mt"""
+
+    NAME = "ARC Finnish"
+    DATASET_PATH = "LumiOpen/arc_challenge_mt"
+    SAMPLE_SPLIT = "test"
+    FEWSHOT_SPLIT = "validation"
+    RESPONSE_TYPE = ResponseType.LOGLIKELIHOODS
+    METRICS = [AccuracyLoglikelihood, AccuracyNormLoglikelihood]
+    SUBJECTS = ["fi"]
+    PERTURBATION_UNMODIFIABLE_WORDS = ["Question"] + get_n_letters(5)
+    LANGUAGE = Language.FIN
+
+    def __init__(self, num_fewshot: int = 0) -> None:
+        super().__init__(num_fewshot)
+
+        self.keys = get_n_letters(5)  # needs to be 5 because there is one sample with 5 answer possibilities
+        self.num_to_letter = {str(i): letter for i, letter in enumerate(self.keys, start=1)}
+
+    def _get_instruction_text(self, item: dict[str, Any]) -> str:
+        return f"Question: {item['question']}\n"
+
+    def _get_fewshot_target_text(self, item: dict[str, Any]) -> str:
+        ground_truth = self._get_ground_truth(item)
+        assert ground_truth is not None
+        return f"{self._get_cue_text(item)}{ground_truth}"
+
+    def _get_cue_text(self, item: dict[str, Any]) -> str:
+        return "Answer:"
+
+    def _get_ground_truth(self, item: dict[str, Any]) -> str | None:
+        answer_key = self.num_to_letter.get(item["answerKey"], item["answerKey"])
+        return f" {item['choices']['text'][self.keys.index(answer_key)]}"
+
+    def _get_possible_completions(self, item: dict[str, Any]) -> list[str] | None:
+        return [f" {choice}" for choice in item["choices"]["text"]]