Merge branch 'master' of github.com:EleutherAI/lm-evaluation-harness into...

Merge branch 'master' of github.com:EleutherAI/lm-evaluation-harness into evaluator-description-option

lm_eval/base.py

@@ -11,7 +11,7 @@ from tqdm import tqdm
 import torch
 import torch.nn.functional as F
 
-from lm_eval.metrics import mean, weighted_perplexity, weighted_mean
+from lm_eval.metrics import mean, weighted_perplexity, weighted_mean, bits_per_byte
 from lm_eval import utils
 from abc import abstractmethod
 
@@ -599,14 +599,14 @@ class PerplexityTask(Task, abc.ABC):
         return {
             "word_perplexity": (loglikelihood, words),
             "byte_perplexity": (loglikelihood, bytes_),
-            "bits_per_byte": (-loglikelihood, self.count_bytes(doc))
+            "bits_per_byte": (loglikelihood, bytes_),
         }
 
     def aggregation(self):
         return {
             "word_perplexity": weighted_perplexity,
             "byte_perplexity": weighted_perplexity,
-            "bits_per_byte": weighted_mean
+            "bits_per_byte": bits_per_byte,
         }
 
     @classmethod

lm_eval/metrics.py

@@ -102,6 +102,9 @@ def weighted_mean(items):
 def weighted_perplexity(items):
     return math.exp(-weighted_mean(items))
 
+def bits_per_byte(items):
+    return -weighted_mean(items) / math.log(2)
+
 
 def bleu(items):
     """The Bilingual Evaluation Understudy Score, or BLEU for short, is a metric

lm_eval/tasks/__init__.py

@@ -46,6 +46,8 @@ from . import wikitext
 from . import lambada_multilingual
 from . import mutual
 from . import truthfulqa
+from . import blimp
+from . import asdiv
 
 ########################################
 # Translation tasks
@@ -134,7 +136,9 @@ TASK_REGISTRY = {
     "squad2": squad.SQuAD2,
     "race": race.RACE,
     # "naturalqs": naturalqs.NaturalQs, # not implemented yet
-    "headqa": headqa.HeadQA,
+    "headqa": headqa.HeadQAEsDeprecated, # for backwards compat - headqa used to default to es
+    "headqa_es": headqa.HeadQAEs,
+    "headqa_en": headqa.HeadQAEn,
     "mathqa": mathqa.MathQA,
     "webqs": webqs.WebQs,
     "wsc273": wsc273.WinogradSchemaChallenge273,
@@ -165,6 +169,7 @@ TASK_REGISTRY = {
     "math_num_theory": hendrycks_math.MathNumberTheory,
     "math_prealgebra": hendrycks_math.MathPrealgebra,
     "math_precalc": hendrycks_math.MathPrecalculus,
+    "math_asdiv": asdiv.Asdiv,
 
     # arithmetic
     "arithmetic_2da": arithmetic.Arithmetic2DPlus,
@@ -219,6 +224,74 @@ TASK_REGISTRY = {
     "pile_wikipedia": pile.PileWikipedia,
     "pile_youtubesubtitles": pile.PileYoutubeSubtitles,
     
+    # BLiMP
+    "blimp_adjunct_island": blimp.BlimpAdjunctIsland,
+    "blimp_anaphor_gender_agreement": blimp.BlimpAnaphorGenderAgreement,
+    "blimp_anaphor_number_agreement": blimp.BlimpAnaphorNumberAgreement,
+    "blimp_animate_subject_passive": blimp.BlimpAnimateSubjectPassive,
+    "blimp_animate_subject_trans": blimp.BlimpAnimateSubjectTrans,
+    "blimp_causative": blimp.BlimpCausative,
+    "blimp_complex_NP_island": blimp.BlimpComplex_NPIsland,
+    "blimp_coordinate_structure_constraint_complex_left_branch": blimp.BlimpCoordinateStructureConstraintComplexLeftBranch,
+    "blimp_coordinate_structure_constraint_object_extraction": blimp.BlimpCoordinateStructureConstraintObjectExtraction,
+    "blimp_determiner_noun_agreement_1": blimp.BlimpDeterminerNounAgreement_1,
+    "blimp_determiner_noun_agreement_2": blimp.BlimpDeterminerNounAgreement_2,
+    "blimp_determiner_noun_agreement_irregular_1": blimp.BlimpDeterminerNounAgreementIrregular_1,
+    "blimp_determiner_noun_agreement_irregular_2": blimp.BlimpDeterminerNounAgreementIrregular_2,
+    "blimp_determiner_noun_agreement_with_adj_2": blimp.BlimpDeterminerNounAgreementWithAdj_2,
+    "blimp_determiner_noun_agreement_with_adj_irregular_1": blimp.BlimpDeterminerNounAgreementWithAdjIrregular_1,
+    "blimp_determiner_noun_agreement_with_adj_irregular_2": blimp.BlimpDeterminerNounAgreementWithAdjIrregular_2,
+    "blimp_determiner_noun_agreement_with_adjective_1": blimp.BlimpDeterminerNounAgreementWithAdjective_1,
+    "blimp_distractor_agreement_relational_noun": blimp.BlimpDistractorAgreementRelationalNoun,
+    "blimp_distractor_agreement_relative_clause": blimp.BlimpDistractorAgreementRelativeClause,
+    "blimp_drop_argument": blimp.BlimpDropArgument,
+    "blimp_ellipsis_n_bar_1": blimp.BlimpEllipsisNBar_1,
+    "blimp_ellipsis_n_bar_2": blimp.BlimpEllipsisNBar_2,
+    "blimp_existential_there_object_raising": blimp.BlimpExistentialThereObjectRaising,
+    "blimp_existential_there_quantifiers_1": blimp.BlimpExistentialThereQuantifiers_1,
+    "blimp_existential_there_quantifiers_2": blimp.BlimpExistentialThereQuantifiers_2,
+    "blimp_existential_there_subject_raising": blimp.BlimpExistentialThereSubjectRaising,
+    "blimp_expletive_it_object_raising": blimp.BlimpExpletiveItObjectRaising,
+    "blimp_inchoative": blimp.BlimpInchoative,
+    "blimp_intransitive": blimp.BlimpIntransitive,
+    "blimp_irregular_past_participle_adjectives": blimp.BlimpIrregularPastParticipleAdjectives,
+    "blimp_irregular_past_participle_verbs": blimp.BlimpIrregularPastParticipleVerbs,
+    "blimp_irregular_plural_subject_verb_agreement_1": blimp.BlimpIrregularPluralSubjectVerbAgreement_1,
+    "blimp_irregular_plural_subject_verb_agreement_2": blimp.BlimpIrregularPluralSubjectVerbAgreement_2,
+    "blimp_left_branch_island_echo_question": blimp.BlimpLeftBranchIslandEchoQuestion,
+    "blimp_left_branch_island_simple_question": blimp.BlimpLeftBranchIslandSimpleQuestion,
+    "blimp_matrix_question_npi_licensor_present": blimp.BlimpMatrixQuestionNpiLicensorPresent,
+    "blimp_npi_present_1": blimp.BlimpNpiPresent_1,
+    "blimp_npi_present_2": blimp.BlimpNpiPresent_2,
+    "blimp_only_npi_licensor_present": blimp.BlimpOnlyNpiLicensorPresent,
+    "blimp_only_npi_scope": blimp.BlimpOnlyNpiScope,
+    "blimp_passive_1": blimp.BlimpPassive_1,
+    "blimp_passive_2": blimp.BlimpPassive_2,
+    "blimp_principle_A_c_command": blimp.BlimpPrinciple_ACCommand,
+    "blimp_principle_A_case_1": blimp.BlimpPrinciple_ACase_1,
+    "blimp_principle_A_case_2": blimp.BlimpPrinciple_ACase_2,
+    "blimp_principle_A_domain_1": blimp.BlimpPrinciple_ADomain_1,
+    "blimp_principle_A_domain_2": blimp.BlimpPrinciple_ADomain_2,
+    "blimp_principle_A_domain_3": blimp.BlimpPrinciple_ADomain_3,
+    "blimp_principle_A_reconstruction": blimp.BlimpPrinciple_AReconstruction,
+    "blimp_regular_plural_subject_verb_agreement_1": blimp.BlimpRegularPluralSubjectVerbAgreement_1,
+    "blimp_regular_plural_subject_verb_agreement_2": blimp.BlimpRegularPluralSubjectVerbAgreement_2,
+    "blimp_sentential_negation_npi_licensor_present": blimp.BlimpSententialNegationNpiLicensorPresent,
+    "blimp_sentential_negation_npi_scope": blimp.BlimpSententialNegationNpiScope,
+    "blimp_sentential_subject_island": blimp.BlimpSententialSubjectIsland,
+    "blimp_superlative_quantifiers_1": blimp.BlimpSuperlativeQuantifiers_1,
+    "blimp_superlative_quantifiers_2": blimp.BlimpSuperlativeQuantifiers_2,
+    "blimp_tough_vs_raising_1": blimp.BlimpToughVsRaising_1,
+    "blimp_tough_vs_raising_2": blimp.BlimpToughVsRaising_2,
+    "blimp_transitive": blimp.BlimpTransitive,
+    "blimp_wh_island": blimp.BlimpWhIsland,
+    "blimp_wh_questions_object_gap": blimp.BlimpWhQuestionsObjectGap,
+    "blimp_wh_questions_subject_gap": blimp.BlimpWhQuestionsSubjectGap,
+    "blimp_wh_questions_subject_gap_long_distance": blimp.BlimpWhQuestionsSubjectGapLongDistance,
+    "blimp_wh_vs_that_no_gap": blimp.BlimpWhVsThatNoGap,
+    "blimp_wh_vs_that_no_gap_long_distance": blimp.BlimpWhVsThatNoGapLongDistance,
+    "blimp_wh_vs_that_with_gap": blimp.BlimpWhVsThatWithGap,
+    "blimp_wh_vs_that_with_gap_long_distance": blimp.BlimpWhVsThatWithGapLongDistance,
 }
 
 

lm_eval/tasks/asdiv.py

+"""
+ASDiv: A Diverse Corpus for Evaluating and Developing English Math Word Problem Solvers
+https://arxiv.org/abs/2106.15772
+
+@misc{miao2021diverse,
+      title={A Diverse Corpus for Evaluating and Developing English Math Word Problem Solvers},
+      author={Shen-Yun Miao and Chao-Chun Liang and Keh-Yih Su},
+      year={2021},
+      eprint={2106.15772},
+      archivePrefix={arXiv},
+      primaryClass={cs.AI}
+}
+"""
+from lm_eval.base import Task
+from pathlib import Path
+from best_download import download_file 
+import xml.etree.ElementTree as ET
+from lm_eval.base import rf
+from lm_eval.metrics import mean,perplexity
+import numpy as np
+from zipfile import ZipFile
+import os 
+
+#currently ignoring formula for answer generation
+
+# given a subset, splits return the docs 
+class Asdiv(Task):
+    VERSION = 0
+    DATASET_PATH = Path("data/asdiv")
+
+    def download(self):
+        if self.DATASET_PATH.exists():
+            return
+        Path.mkdir(self.DATASET_PATH)
+        url = "https://github.com/chaochun/nlu-asdiv-dataset/archive/55790e5270bb91ccfa5053194b25732534696b50.zip"
+        checksum = "8f1fe4f6d5f170ec1e24ab78c244153c14c568b1bb2b1dad0324e71f37939a2d"
+        zip_path = self.DATASET_PATH / "55790e5270bb91ccfa5053194b25732534696b50.zip"
+        download_file(url, str(zip_path), checksum)
+        with ZipFile(zip_path, "r") as zip:
+            zip.extractall(self.DATASET_PATH)
+        os.remove(zip_path)
+
+    def _convert_standard(self, problem):
+        #TODO: include solution-type and formula
+        out_doc = {
+            "question" : problem.find('Question').text,
+            "body" : problem.find('Body').text,
+            "answer": problem.find('Answer').text
+        }
+        return out_doc
+
+    def load_docs(self, textfilename, tfds=False):
+        tree = ET.parse(textfilename)
+        root = tree.getroot()
+        for pid, problem in enumerate(root.iter('Problem')):
+            out_doc = self._convert_standard(problem)
+            yield out_doc
+
+    def has_training_docs(self):
+        return False
+    
+    def has_validation_docs(self):
+        return True
+
+    def has_test_docs(self):
+        return False
+
+    def training_docs(self):
+        raise NotImplementedError("This dataset has no training docs")
+
+    def test_docs(self):
+        raise NotImplementedError("This dataset has no test docs")
+
+    def validation_docs(self):
+        data_xml_path = self.DATASET_PATH / "nlu-asdiv-dataset-55790e5270bb91ccfa5053194b25732534696b50/dataset/ASDiv.xml"
+        return self.load_docs(data_xml_path)
+
+    def fewshot_context(self, doc, num_fewshot, provide_description, rnd):
+        assert num_fewshot == 0, "ASDiv is intended only for the zero-shot setting."
+        return super().fewshot_context(doc, num_fewshot, provide_description, rnd)
+
+    
+    def fewshot_description(self):
+        # TODO: add solution-type and formula
+        desc = "information containing the context of the question\nQuestion: Text of a question.\nAnswer: Answer to the question, based on the passage.\n"
+        return desc
+
+    def doc_to_text(self, doc):
+        # TODO: add solution-type
+        return doc['body'] + '\n' + 'Question:' + doc['question'] + '\n' + 'Answer:'
+
+    def doc_to_target(self, doc):
+        # TODO: add formula
+
+        answer = doc['answer'].split(' (')[0]
+        return " " + answer
+
+    def construct_requests(self, doc, ctx):
+        ll, is_greedy = rf.loglikelihood(ctx, self.doc_to_target(doc))
+        return ll, is_greedy
+    
+    def process_results(self, doc, results):
+        ll, is_greedy = results
+
+        return {
+            'acc': int(is_greedy)
+        }
+        
+    def aggregation(self):
+        return {
+            'acc': mean
+        }
+
+    def higher_is_better(self):
+        return {
+            'acc': True
+        }
+

lm_eval/tasks/blimp.py

+"""
+BLiMP: A Benchmark of Linguistic Minimal Pairs for English
+https://arxiv.org/abs/1912.00582
+
+@article{warstadt2019blimp,
+  title={BLiMP: A Benchmark of Linguistic Minimal Pairs for English},
+  author={Warstadt, Alex and Parrish, Alicia and Liu, Haokun and Mohananey, Anhad and Peng, Wei, and Wang, Sheng-Fu and Bowman, Samuel R},
+  journal={arXiv preprint arXiv:1912.00582},
+  year={2019}
+}
+"""
+
+from lm_eval.base import rf
+from lm_eval.metrics import mean
+from .common import HFTask
+
+
+class BlimpTask(HFTask):
+    VERSION = 0
+    DATASET_PATH = "blimp"
+
+    def download(self):
+        super().download()
+
+        # The HF dataset only contains a "train" dataset, but the harness expects a "validation"
+        # dataset. Let's use the training dataset, on the assumption that the model wasn't actually
+        # trained on this data.
+
+        self.data["validation"] = self.data["train"]
+        del self.data["train"]
+
+    def fewshot_context(self, doc, num_fewshot, provide_description, rnd):
+        assert num_fewshot == 0
+        assert not provide_description
+        return ""
+
+    def doc_to_text(self, doc):
+        # this method is invoked by tests only
+        return ""
+
+    def doc_to_target(self, doc):
+        # this method is invoked by tests only
+        return ""
+
+    def construct_requests(self, doc, ctx):
+        assert not ctx
+
+        # Calculate the loglikelihood for the good and the bad sentence.
+        # Note that loglikelihood translates the "" prefix to the "<|endoftext|>" token
+        return [
+            rf.loglikelihood("", doc["sentence_good"]),
+            rf.loglikelihood("", doc["sentence_bad"]),
+        ]
+
+    def process_results(self, doc, results):
+        likelihood1, likelihood2 = results
+
+        # the model got this case right iff the good sentence scored higher than the bad sentence
+        acc = 1.0 if likelihood1 > likelihood2 else 0.0
+
+        return {
+            "acc": acc,
+        }
+
+    def higher_is_better(self):
+        return {
+            "acc": True,
+        }
+
+    def aggregation(self):
+        return {
+            "acc": mean,
+        }
+
+
+class BlimpAdjunctIsland(BlimpTask):
+    DATASET_NAME = "adjunct_island"
+
+
+class BlimpAnaphorGenderAgreement(BlimpTask):
+    DATASET_NAME = "anaphor_gender_agreement"
+
+
+class BlimpAnaphorNumberAgreement(BlimpTask):
+    DATASET_NAME = "anaphor_number_agreement"
+
+
+class BlimpAnimateSubjectPassive(BlimpTask):
+    DATASET_NAME = "animate_subject_passive"
+
+
+class BlimpAnimateSubjectTrans(BlimpTask):
+    DATASET_NAME = "animate_subject_trans"
+
+
+class BlimpCausative(BlimpTask):
+    DATASET_NAME = "causative"
+
+
+class BlimpComplex_NPIsland(BlimpTask):
+    DATASET_NAME = "complex_NP_island"
+
+
+class BlimpCoordinateStructureConstraintComplexLeftBranch(BlimpTask):
+    DATASET_NAME = "coordinate_structure_constraint_complex_left_branch"
+
+
+class BlimpCoordinateStructureConstraintObjectExtraction(BlimpTask):
+    DATASET_NAME = "coordinate_structure_constraint_object_extraction"
+
+
+class BlimpDeterminerNounAgreement_1(BlimpTask):
+    DATASET_NAME = "determiner_noun_agreement_1"
+
+
+class BlimpDeterminerNounAgreement_2(BlimpTask):
+    DATASET_NAME = "determiner_noun_agreement_2"
+
+
+class BlimpDeterminerNounAgreementIrregular_1(BlimpTask):
+    DATASET_NAME = "determiner_noun_agreement_irregular_1"
+
+
+class BlimpDeterminerNounAgreementIrregular_2(BlimpTask):
+    DATASET_NAME = "determiner_noun_agreement_irregular_2"
+
+
+class BlimpDeterminerNounAgreementWithAdj_2(BlimpTask):
+    DATASET_NAME = "determiner_noun_agreement_with_adj_2"
+
+
+class BlimpDeterminerNounAgreementWithAdjIrregular_1(BlimpTask):
+    DATASET_NAME = "determiner_noun_agreement_with_adj_irregular_1"
+
+
+class BlimpDeterminerNounAgreementWithAdjIrregular_2(BlimpTask):
+    DATASET_NAME = "determiner_noun_agreement_with_adj_irregular_2"
+
+
+class BlimpDeterminerNounAgreementWithAdjective_1(BlimpTask):
+    DATASET_NAME = "determiner_noun_agreement_with_adjective_1"
+
+
+class BlimpDistractorAgreementRelationalNoun(BlimpTask):
+    DATASET_NAME = "distractor_agreement_relational_noun"
+
+
+class BlimpDistractorAgreementRelativeClause(BlimpTask):
+    DATASET_NAME = "distractor_agreement_relative_clause"
+
+
+class BlimpDropArgument(BlimpTask):
+    DATASET_NAME = "drop_argument"
+
+
+class BlimpEllipsisNBar_1(BlimpTask):
+    DATASET_NAME = "ellipsis_n_bar_1"
+
+
+class BlimpEllipsisNBar_2(BlimpTask):
+    DATASET_NAME = "ellipsis_n_bar_2"
+
+
+class BlimpExistentialThereObjectRaising(BlimpTask):
+    DATASET_NAME = "existential_there_object_raising"
+
+
+class BlimpExistentialThereQuantifiers_1(BlimpTask):
+    DATASET_NAME = "existential_there_quantifiers_1"
+
+
+class BlimpExistentialThereQuantifiers_2(BlimpTask):
+    DATASET_NAME = "existential_there_quantifiers_2"
+
+
+class BlimpExistentialThereSubjectRaising(BlimpTask):
+    DATASET_NAME = "existential_there_subject_raising"
+
+
+class BlimpExpletiveItObjectRaising(BlimpTask):
+    DATASET_NAME = "expletive_it_object_raising"
+
+
+class BlimpInchoative(BlimpTask):
+    DATASET_NAME = "inchoative"
+
+
+class BlimpIntransitive(BlimpTask):
+    DATASET_NAME = "intransitive"
+
+
+class BlimpIrregularPastParticipleAdjectives(BlimpTask):
+    DATASET_NAME = "irregular_past_participle_adjectives"
+
+
+class BlimpIrregularPastParticipleVerbs(BlimpTask):
+    DATASET_NAME = "irregular_past_participle_verbs"
+
+
+class BlimpIrregularPluralSubjectVerbAgreement_1(BlimpTask):
+    DATASET_NAME = "irregular_plural_subject_verb_agreement_1"
+
+
+class BlimpIrregularPluralSubjectVerbAgreement_2(BlimpTask):
+    DATASET_NAME = "irregular_plural_subject_verb_agreement_2"
+
+
+class BlimpLeftBranchIslandEchoQuestion(BlimpTask):
+    DATASET_NAME = "left_branch_island_echo_question"
+
+
+class BlimpLeftBranchIslandSimpleQuestion(BlimpTask):
+    DATASET_NAME = "left_branch_island_simple_question"
+
+
+class BlimpMatrixQuestionNpiLicensorPresent(BlimpTask):
+    DATASET_NAME = "matrix_question_npi_licensor_present"
+
+
+class BlimpNpiPresent_1(BlimpTask):
+    DATASET_NAME = "npi_present_1"
+
+
+class BlimpNpiPresent_2(BlimpTask):
+    DATASET_NAME = "npi_present_2"
+
+
+class BlimpOnlyNpiLicensorPresent(BlimpTask):
+    DATASET_NAME = "only_npi_licensor_present"
+
+
+class BlimpOnlyNpiScope(BlimpTask):
+    DATASET_NAME = "only_npi_scope"
+
+
+class BlimpPassive_1(BlimpTask):
+    DATASET_NAME = "passive_1"
+
+
+class BlimpPassive_2(BlimpTask):
+    DATASET_NAME = "passive_2"
+
+
+class BlimpPrinciple_ACCommand(BlimpTask):
+    DATASET_NAME = "principle_A_c_command"
+
+
+class BlimpPrinciple_ACase_1(BlimpTask):
+    DATASET_NAME = "principle_A_case_1"
+
+
+class BlimpPrinciple_ACase_2(BlimpTask):
+    DATASET_NAME = "principle_A_case_2"
+
+
+class BlimpPrinciple_ADomain_1(BlimpTask):
+    DATASET_NAME = "principle_A_domain_1"
+
+
+class BlimpPrinciple_ADomain_2(BlimpTask):
+    DATASET_NAME = "principle_A_domain_2"
+
+
+class BlimpPrinciple_ADomain_3(BlimpTask):
+    DATASET_NAME = "principle_A_domain_3"
+
+
+class BlimpPrinciple_AReconstruction(BlimpTask):
+    DATASET_NAME = "principle_A_reconstruction"
+
+
+class BlimpRegularPluralSubjectVerbAgreement_1(BlimpTask):
+    DATASET_NAME = "regular_plural_subject_verb_agreement_1"
+
+
+class BlimpRegularPluralSubjectVerbAgreement_2(BlimpTask):
+    DATASET_NAME = "regular_plural_subject_verb_agreement_2"
+
+
+class BlimpSententialNegationNpiLicensorPresent(BlimpTask):
+    DATASET_NAME = "sentential_negation_npi_licensor_present"
+
+
+class BlimpSententialNegationNpiScope(BlimpTask):
+    DATASET_NAME = "sentential_negation_npi_scope"
+
+
+class BlimpSententialSubjectIsland(BlimpTask):
+    DATASET_NAME = "sentential_subject_island"
+
+
+class BlimpSuperlativeQuantifiers_1(BlimpTask):
+    DATASET_NAME = "superlative_quantifiers_1"
+
+
+class BlimpSuperlativeQuantifiers_2(BlimpTask):
+    DATASET_NAME = "superlative_quantifiers_2"
+
+
+class BlimpToughVsRaising_1(BlimpTask):
+    DATASET_NAME = "tough_vs_raising_1"
+
+
+class BlimpToughVsRaising_2(BlimpTask):
+    DATASET_NAME = "tough_vs_raising_2"
+
+
+class BlimpTransitive(BlimpTask):
+    DATASET_NAME = "transitive"
+
+
+class BlimpWhIsland(BlimpTask):
+    DATASET_NAME = "wh_island"
+
+
+class BlimpWhQuestionsObjectGap(BlimpTask):
+    DATASET_NAME = "wh_questions_object_gap"
+
+
+class BlimpWhQuestionsSubjectGap(BlimpTask):
+    DATASET_NAME = "wh_questions_subject_gap"
+
+
+class BlimpWhQuestionsSubjectGapLongDistance(BlimpTask):
+    DATASET_NAME = "wh_questions_subject_gap_long_distance"
+
+
+class BlimpWhVsThatNoGap(BlimpTask):
+    DATASET_NAME = "wh_vs_that_no_gap"
+
+
+class BlimpWhVsThatNoGapLongDistance(BlimpTask):
+    DATASET_NAME = "wh_vs_that_no_gap_long_distance"
+
+
+class BlimpWhVsThatWithGap(BlimpTask):
+    DATASET_NAME = "wh_vs_that_with_gap"
+
+
+class BlimpWhVsThatWithGapLongDistance(BlimpTask):
+    DATASET_NAME = "wh_vs_that_with_gap_long_distance"

lm_eval/tasks/glue.py

@@ -220,7 +220,7 @@ class QNLI(HFTask):
 
 
 class WNLI(HFTask):
-    VERSION = 0
+    VERSION = 1
     DATASET_PATH = "glue"
     DATASET_NAME = "wnli"
 
@@ -234,26 +234,25 @@ class WNLI(HFTask):
         return False
 
     def doc_to_text(self, doc):
-        return "{}\nQuestion: {} True, False or Neither?\nAnswer:".format(
+        return "{}\nQuestion: {} True or False?\nAnswer:".format(
             doc["sentence1"],
             doc["sentence2"],
         )
 
     def doc_to_target(self, doc):
         # True = entailment
-        # False = contradiction
-        # Neither = neutral
-        return " {}".format({0: "True", 1: "Neither", 2: "False"}[doc["label"]])
+        # False = not_entailment
+        return " {}".format({0: "False", 1: "True"}[doc["label"]])
 
     def construct_requests(self, doc, ctx):
         ll_true, _ = rf.loglikelihood(ctx, " True")
-        ll_neither, _ = rf.loglikelihood(ctx, " Neither")
         ll_false, _ = rf.loglikelihood(ctx, " False")
-        return ll_true, ll_neither, ll_false
+        return ll_true, ll_false
 
     def process_results(self, doc, results):
+        ll_true, ll_false = results
+        pred = ll_true > ll_false
         gold = doc["label"]
-        pred = np.argmax(results)
         return {
             "acc": pred == gold
         }

lm_eval/tasks/headqa.py

@@ -2,10 +2,9 @@ from . common import HFTask
 from lm_eval.base import MultipleChoiceTask
 
 
-class HeadQA(HFTask, MultipleChoiceTask):
+class HeadQABase(HFTask, MultipleChoiceTask):
     VERSION = 0
     DATASET_PATH = "head_qa"
-    DATASET_NAME = None
 
     def has_training_docs(self):
         return True
@@ -27,3 +26,15 @@ class HeadQA(HFTask, MultipleChoiceTask):
 
     def doc_to_text(self, doc):
         return doc["query"]
+
+class HeadQAEn(HeadQABase):
+    DATASET_NAME = "en"
+
+class HeadQAEs(HeadQABase):
+    DATASET_NAME = "es"
+
+# for backwards compatibility
+class HeadQAEsDeprecated(HeadQABase):
+    DATASET_NAME = "es"
+
+    print("WARNING: headqa is deprecated. Please use headqa_es or headqa_en instead. See https://github.com/EleutherAI/lm-evaluation-harness/pull/240 for more info.")
\ No newline at end of file

lm_eval/tasks/pile.py

@@ -10,7 +10,7 @@ from best_download import download_file
 
 
 class PilePerplexityTask(PerplexityTask, abc.ABC):
-    VERSION = 0
+    VERSION = 1
 
     PILE_SET_NAME = None
     VAL_PATH = 'data/pile/val.jsonl.zst'
@@ -18,9 +18,12 @@ class PilePerplexityTask(PerplexityTask, abc.ABC):
 
     def download(self):
         # TODO: separate pile val/test out by component so we don't have to scan the entire file once per set
-        os.makedirs("data/pile/", exist_ok=True)
-        download_file("https://the-eye.eu/public/AI/pile/val.jsonl.zst", self.VAL_PATH, "264c875d8bbd355d8daa9d032b75fd8fb91606218bb84dd1155b203fcd5fab92")
-        download_file("https://the-eye.eu/public/AI/pile/test.jsonl.zst", self.TEST_PATH, "0bb28c52d0b5596d389bf179ce2d43bf7f7ffae76b0d2d20b180c97f62e0975e")
+
+        if not os.path.exists("data/pile/test.jsonl.zst"):
+            # todo use new best_download fallback api
+            os.makedirs("data/pile/", exist_ok=True)
+            download_file("http://eaidata.bmk.sh/data/pile/val.jsonl.zst", self.VAL_PATH, "264c875d8bbd355d8daa9d032b75fd8fb91606218bb84dd1155b203fcd5fab92")
+            download_file("http://eaidata.bmk.sh/data/pile/test.jsonl.zst", self.TEST_PATH, "0bb28c52d0b5596d389bf179ce2d43bf7f7ffae76b0d2d20b180c97f62e0975e")
 
     def validation_docs(self):
         rdr = lm_dataformat.Reader(self.VAL_PATH)

lm_eval/tasks/superglue.py

@@ -13,7 +13,7 @@ from ..utils import general_detokenize
 
 
 class BoolQ(HFTask):
-    VERSION = 0
+    VERSION = 1
     DATASET_PATH = "super_glue"
     DATASET_NAME = "boolq"
 
@@ -27,7 +27,7 @@ class BoolQ(HFTask):
         return False
 
     def doc_to_text(self, doc):
-        return f"{doc['passage']}\nQuestion: {doc['question']}\nAnswer:"
+        return f"{doc['passage']}\nQuestion: {doc['question']}?\nAnswer:"
     
     def doc_to_target(self, doc):
         return " " + yesno(doc['label']) 

lm_eval/tasks/wikitext.py

@@ -41,7 +41,7 @@ def wikitext_detokenizer(string):
 
 
 class WikiText(PerplexityTask):
-    VERSION = 0
+    VERSION = 1
 
     def download(self):
         if not os.path.exists('data/wikitext/wikitext-2-raw/wiki.valid.raw'):

tests/test_tasks.py

@@ -32,7 +32,7 @@ def test_basic_interface(taskname, task_class):
 
     limit = None
 
-    if taskname in ["triviaqa"]:
+    if taskname in ["triviaqa"] or taskname.startswith("pile_"):
         limit = 10000
     if task.has_validation_docs():
         arr = list(islice(task.validation_docs(), limit))

tests/testdata/boolq-v1-loglikelihood

+6577e0d88572772ef08e64f624c0e3df0953286ae1f118ccef15623b59ffeabf
\ No newline at end of file

tests/testdata/boolq-v1-res.json

+{"results": {"boolq": {"acc": 0.5048929663608562, "acc_stderr": 0.00874463623355505}}, "versions": {"boolq": 1}}
\ No newline at end of file

tests/testdata/headqa_en-v0-loglikelihood

+09da45119b12a0144e3081f8fb790c2a22af7b9c3aac42f54423d348a711fbf5
\ No newline at end of file

tests/testdata/headqa_en-v0-res.json

+{"results": {"headqa_en": {"acc": 0.23559445660102116, "acc_norm": 0.2447118891320204, "acc_norm_stderr": 0.008211629406841468, "acc_stderr": 0.008105688874297972}}, "versions": {"headqa_en": 0}}
\ No newline at end of file

tests/testdata/headqa_es-v0-loglikelihood

+767ca34d9714edd9fb030ddbcc35a64e5180d1e247b0cb557fbb22fdf971ad1f
\ No newline at end of file

tests/testdata/headqa_es-v0-res.json

+{"results": {"headqa_es": {"acc": 0.23559445660102116, "acc_norm": 0.25018234865062, "acc_norm_stderr": 0.008272783230806014, "acc_stderr": 0.008105688874297972}}, "versions": {"headqa_es": 0}}
\ No newline at end of file

tests/testdata/pile_arxiv-v1-loglikelihood_rolling

+814f9954e44368559602c00f7e85fa3971acdfd0315f508ec7df6318a79c55ec
\ No newline at end of file

tests/testdata/pile_arxiv-v1-res.json

+{"results": {"pile_arxiv": {"bits_per_byte": 1.55095665856779e-05, "byte_perplexity": 1.0000107504701365, "word_perplexity": 1.0000819333090385}}, "versions": {"pile_arxiv": 1}}
\ No newline at end of file