Source code for eval_framework.tasks.benchmarks.sciq

import hashlib
import random
from typing import Any

from eval_framework.metrics.loglikelihood.accuracy_loglikelihood import (
    AccuracyLoglikelihood,
    AccuracyNormLoglikelihood,
)
from eval_framework.metrics.loglikelihood.bits_per_byte import BitsPerByteLoglikelihood
from eval_framework.metrics.loglikelihood.confidence_weighted_accuracy import ConfidenceWeightedAccuracy
from eval_framework.metrics.loglikelihood.dcs import DistributionalCorrectnessScore
from eval_framework.metrics.loglikelihood.ternary import TernaryScore
from eval_framework.tasks.base import NO_SUBJECT, BaseTask, Language, ResponseType
from eval_framework.tasks.utils import get_n_letters


def _shuffled_choices_and_correct_index(item: dict[str, Any]) -> tuple[list[str], int]:
    """Return (shuffled_choices, correct_index) with deterministic shuffle from item content."""
    choices = [
        item["distractor1"],
        item["distractor2"],
        item["distractor3"],
        item["correct_answer"],
    ]
    seed = int(hashlib.sha256((item["question"] + item["correct_answer"]).encode()).hexdigest(), 16)
    rng = random.Random(seed)
    order = list(range(4))
    rng.shuffle(order)
    shuffled = [choices[i] for i in order]
    correct_index = order.index(3)  # 3 = index of correct_answer in original list
    return shuffled, correct_index


[docs] class SCIQ(BaseTask[str]): """SciQ dataset: https://huggingface.co/datasets/allenai/sciq""" NAME = "SciQ" DATASET_PATH = "allenai/sciq" SAMPLE_SPLIT = "validation" # 1000 examples (same split as lm-eval) FEWSHOT_SPLIT = "test" # 1000 examples RESPONSE_TYPE = ResponseType.LOGLIKELIHOODS METRICS = [AccuracyLoglikelihood, AccuracyNormLoglikelihood, BitsPerByteLoglikelihood] SUBJECTS = [NO_SUBJECT] PERTURBATION_UNMODIFIABLE_WORDS = ["Question"] LANGUAGE = Language.ENG def _get_instruction_text(self, item: dict[str, Any]) -> str: return f"Question: {item['question']}\n" def _get_ground_truth(self, item: dict[str, Any]) -> str | None: return f" {item['correct_answer']}" 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_possible_completions(self, item: dict[str, Any]) -> list[str] | None: shuffled, _ = _shuffled_choices_and_correct_index(item) return [f" {choice}" for choice in shuffled]
[docs] class SCIQ_OLMES(SCIQ): """ SciQ with OLMES-style prompt: options shown with space-prefixed labels (" A.", " B.", " C.", " D."); loglikelihood over " A"/" B"/" C"/" D". Answer choices are deterministically shuffled per example. """ NAME = "SciQ_OLMES" SAMPLE_SPLIT = "train" # Use train split (largest) to best match OLMES, which evaluates all splits FEWSHOT_SPLIT = "train" def __init__(self, num_fewshot: int = 0) -> None: super().__init__(num_fewshot) self.keys = get_n_letters(4) def _get_instruction_text(self, item: dict[str, Any]) -> str: question = item["question"] shuffled, _ = _shuffled_choices_and_correct_index(item) options = "\n".join(f" {key}. {choice}" for key, choice in zip(self.keys, shuffled)) return f"Question: {question}\n{options}\n" def _get_ground_truth(self, item: dict[str, Any]) -> str | None: _, correct_index = _shuffled_choices_and_correct_index(item) return f" {self.keys[correct_index]}" def _get_possible_completions(self, item: dict[str, Any]) -> list[str] | None: return [f" {key}" for key in self.keys]
[docs] class SCIQ_IDK(SCIQ): NAME = "SciQ_IDK" METRICS = [ AccuracyLoglikelihood, AccuracyNormLoglikelihood, ConfidenceWeightedAccuracy, DistributionalCorrectnessScore, TernaryScore, ] def _get_initial_prompt_text(self, item: dict[str, Any]) -> str: return ( "Answer only if you are confident, since mistakes may be penalised, while correct answers receive points. " "It is acceptable to answer with 'don't know' if you are unsure, and you will receive 0 points." ) def _get_possible_completions(self, item: dict[str, Any]) -> list[str] | None: completions = super()._get_possible_completions(item) return (completions or []) + [" don't know"]
[docs] class SCIQEvalHarness(SCIQ): """Based on https://github.com/EleutherAI/lm-evaluation-harness/blob/main/lm_eval/tasks/sciq/sciq.yaml#L8 In the Eval Harness implementation, the instruction text includes a context passage. This passage often contains the answer, reducing the benchmark to a straightforward copy-and-paste task. """ NAME = "SciQ Eval Harness" DATASET_PATH = "allenai/sciq" SAMPLE_SPLIT = "validation" # 1000 examples (same split as lm-eval) FEWSHOT_SPLIT = "test" # 1000 examples RESPONSE_TYPE = ResponseType.LOGLIKELIHOODS METRICS = [AccuracyLoglikelihood, AccuracyNormLoglikelihood] SUBJECTS = [NO_SUBJECT] PERTURBATION_UNMODIFIABLE_WORDS = ["Question"] LANGUAGE = Language.ENG def _get_instruction_text(self, item: dict[str, Any]) -> str: return f"{item['support'].lstrip()}\nQuestion: {item['question']}\n"
[docs] class SCIQEvalHarness_IDK(SCIQEvalHarness): NAME = "SciQ Eval Harness_IDK" METRICS = [ AccuracyLoglikelihood, AccuracyNormLoglikelihood, ConfidenceWeightedAccuracy, DistributionalCorrectnessScore, TernaryScore, ] def _get_initial_prompt_text(self, item: dict[str, Any]) -> str: return ( "Answer only if you are confident, since mistakes may be penalised, while correct answers receive points. " "It is acceptable to answer with 'don't know' if you are unsure, and you will receive 0 points." ) def _get_possible_completions(self, item: dict[str, Any]) -> list[str] | None: completions = super()._get_possible_completions(item) return (completions or []) + [" don't know"]