Merge pull request #316 from jon-tow/master

Revert "Merge branch 'master' into master"

lm_eval/tasks/gem_xsum.py

-
-"""
-Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization
-https://arxiv.org/pdf/1808.08745.pdf
-
-The dataset is for the task of abstractive summarization in its extreme form, its about summarizing a document in a single sentence. It introduces extreme summarization, a new single-document summarization task which does not favor extractive strategies and calls for an abstractive modeling approach. The idea is to create a short, one-sentence news summary answering the question "What is the article about?". 
-
-This particularly uses the dataset that is part of the GEM benchmark
-Homepage: https://github.com/EdinburghNLP/XSum
-The GEM Benchmark: Natural Language Generation, its Evaluation and Metrics
-https://arxiv.org/pdf/2102.01672v3.pdf
-Write a Short Description of the task.
-Homepage: https://gem-benchmark.com/data_cards/XSum
-"""
-from lm_eval.base import PromptSourceTask
-from lm_eval.base import Task, rf
-
-
-_CITATION = """
-@InProceedings{xsum-emnlp,
-  author =      "Shashi Narayan and Shay B. Cohen and Mirella Lapata",
-  title =       "Don't Give Me the Details, Just the Summary! {T}opic-Aware Convolutional Neural Networks for Extreme Summarization",
-  booktitle =   "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing ",
-  year =        "2018",
-  address =     "Brussels, Belgium",
-}
-"""
-
-
-
-class GEMXSUMBase(PromptSourceTask):
-    VERSION = 0
-    DATASET_PATH = "GEM/xsum"
-    DATASET_NAME = None
-    SPLIT = None
-
-    def has_training_docs(self):
-        return True
-
-    def has_validation_docs(self):
-        return True
-
-    def has_test_docs(self):
-        return True
-
-    def training_docs(self):
-        if self.has_training_docs():
-            # We cache training documents in `self._training_docs` for faster
-            # few-shot processing. If the data is too large to fit in memory,
-            # return the training data as a generator instead of a list.
-            if self._training_docs is None:
-                self._training_docs = list(self.dataset["train"])
-            return self._training_docs
-
-    def validation_docs(self):
-        if self.has_validation_docs():
-            return self.dataset["validation"]
-
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.dataset["test"]
-
-    def max_generation_length(self):
-            return 64
-
-class GEMXSUM(GEMXSUMBase):
-    '''this is for train/validation/test'''
-    SPLIT = ''
-
-class GEMXSUMChallgeSample(GEMXSUMBase):
-    '''this is for challenge_train_sample/challenge_validation_sample'''
-    SPLIT = 'challenge_sample'
-
-    def has_test_docs(self):
-        return False
-        
-    def training_docs(self):
-        if self.has_training_docs():
-            # We cache training documents in `self._training_docs` for faster
-            # few-shot processing. If the data is too large to fit in memory,
-            # return the training data as a generator instead of a list.
-            if self._training_docs is None:
-                self._training_docs = list(self.dataset["challenge_train_sample"])
-            return self._training_docs
-
-    def validation_docs(self):
-        if self.has_validation_docs():
-            return self.dataset["challenge_validation_sample"]
-
-class GEMXSUMChallgeTestBacktranslation(GEMXSUMBase):
-    '''this is for challenge_test_backtranslation'''
-    SPLIT = 'challenge_test_backtranslation'
-
-    def has_training_docs(self):
-        return False
-
-    def has_validation_docs(self):
-        return False
-
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.dataset[self.SPLIT]
-
-class GEMXSUMChallgeTestBFP02(GEMXSUMBase):
-    '''this is for challenge_test_bfp_02'''
-    SPLIT = 'challenge_test_bfp_02'
-
-    def has_training_docs(self):
-        return False
-
-    def has_validation_docs(self):
-        return False
-
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.dataset[self.SPLIT]
-
-class GEMXSUMChallgeTestBFP05(GEMXSUMBase):
-    '''this is for challenge_test_bfp_05'''
-    SPLIT = 'challenge_test_bfp_05'
-
-    def has_training_docs(self):
-        return False
-
-    def has_validation_docs(self):
-        return False
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.dataset[self.SPLIT]
-
-class GEMXSUMChallgeTestNopunc(GEMXSUMBase):
-    '''this is for challenge_test_nopunc'''
-    SPLIT = 'challenge_test_nopunc'
-
-    def has_training_docs(self):
-        return False
-
-    def has_validation_docs(self):
-        return False
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.dataset[self.SPLIT]
-
-class GEMXSUMChallgeTestCovid(GEMXSUMBase):
-    '''this is for challenge_test_covid'''
-    SPLIT = 'challenge_test_covid'
-
-    def has_training_docs(self):
-        return False
-
-    def has_validation_docs(self):
-        return False
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.dataset[self.SPLIT]
\ No newline at end of file

lm_eval/tasks/glue.py

@@ -14,7 +14,7 @@ respect to a wide range of linguistic phenomena found in natural language.
 Homepage: https://gluebenchmark.com/
 """
 import numpy as np
-from lm_eval.base import PromptSourceTask, rf, Task
+from lm_eval.base import rf, Task
 from lm_eval.metrics import mean, matthews_corrcoef, f1_score, yesno
 from lm_eval.utils import general_detokenize
 
@@ -45,7 +45,7 @@ _CITATION = """
 # Single-Sentence Tasks
 
 
-class CoLA(PromptSourceTask):
+class CoLA(Task):
     VERSION = 0
     DATASET_PATH = "glue"
     DATASET_NAME = "cola"
@@ -67,20 +67,37 @@ class CoLA(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
-    # def process_results(self, doc, results):
-    #     answer_choices_list = self.prompt.get_answer_choices_list(doc)
-    #     pred = np.argmax(results)
-    #     target = answer_choices_list.index(self.doc_to_target(doc).strip())
-    #     return {"mcc": (target, pred)}
+    def doc_to_text(self, doc):
+        return "{}\nQuestion: Does this sentence make sense?\nAnswer:".format(doc["sentence"])
 
-    # def higher_is_better(self):
-    #     return {"mcc": True}
+    def doc_to_target(self, doc):
+        return " {}".format({1: "yes", 0: "no"}[doc["label"]])
 
-    # def aggregation(self):
-    #     return {"mcc": matthews_corrcoef}
+    def construct_requests(self, doc, ctx):
+        ll_true, _ = rf.loglikelihood(ctx, " yes")
+        ll_false, _ = rf.loglikelihood(ctx, " no")
+        return ll_true, ll_false
 
+    def process_results(self, doc, results):
+        ll_true, ll_false = results
+        pred = ll_true > ll_false
+        gold = doc["label"]
+        return {
+            "mcc": (gold, pred)
+        }
 
-class SST(PromptSourceTask):
+    def higher_is_better(self):
+        return {
+            "mcc": True
+        }
+
+    def aggregation(self):
+        return {
+            "mcc": matthews_corrcoef
+        }
+
+
+class SST(Task):
     VERSION = 0
     DATASET_PATH = "glue"
     DATASET_NAME = "sst2"
@@ -102,11 +119,42 @@ class SST(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        return "{}\nQuestion: Is this sentence positive or negative?\nAnswer:".format(
+            general_detokenize(doc["sentence"]),
+        )
+
+    def doc_to_target(self, doc):
+        return " {}".format({1: "positive", 0: "negative"}[doc["label"]])
+
+    def construct_requests(self, doc, ctx):
+        ll_positive, _ = rf.loglikelihood(ctx, " positive")
+        ll_negative, _ = rf.loglikelihood(ctx, " negative")
+        return ll_positive, ll_negative
+
+    def process_results(self, doc, results):
+        ll_positive, ll_negative = results
+        pred = ll_positive > ll_negative
+        gold = doc["label"]
+        return {
+            "acc": pred == gold
+        }
+
+    def higher_is_better(self):
+        return {
+            "acc": True
+        }
+
+    def aggregation(self):
+        return {
+            "acc": mean
+        }
+
 
 # Inference Tasks
 
 
-class MNLI(PromptSourceTask):
+class MNLI(Task):
     VERSION = 0
     DATASET_PATH = "glue"
     DATASET_NAME = "mnli"
@@ -133,6 +181,41 @@ class MNLI(PromptSourceTask):
         if self.has_test_docs():
             return self.dataset["test_matched"]
 
+    def doc_to_text(self, doc):
+        return "{}\nQuestion: {} True, False or Neither?\nAnswer:".format(
+            doc["premise"],
+            doc["hypothesis"].strip() + ('' if doc["hypothesis"].strip().endswith('.') else '.'),
+        )
+
+    def doc_to_target(self, doc):
+        # True = entailment
+        # False = contradiction
+        # Neither = neutral
+        return " {}".format({0: "True", 1: "Neither", 2: "False"}[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
+
+    def process_results(self, doc, results):
+        gold = doc["label"]
+        pred = np.argmax(results)
+        return {
+            "acc": pred == gold
+        }
+
+    def higher_is_better(self):
+        return {
+            "acc": True
+        }
+
+    def aggregation(self):
+        return {
+            "acc": mean
+        }
+
 
 class MNLIMismatched(MNLI):
     VERSION = 0
@@ -146,7 +229,7 @@ class MNLIMismatched(MNLI):
             return self.dataset["test_mismatched"]
 
 
-class QNLI(PromptSourceTask):
+class QNLI(Task):
     VERSION = 0
     DATASET_PATH = "glue"
     DATASET_NAME = "qnli"
@@ -168,8 +251,42 @@ class QNLI(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        return "{}\n{}\nQuestion: Does this response answer the question?\nAnswer:".format(
+            doc["question"],
+            doc["sentence"],
+        )
+
+    def doc_to_target(self, doc):
+        # True = entailment
+        # False = not entailment
+        return " {}".format({0: "yes", 1: "no"}[doc["label"]])
+
+    def construct_requests(self, doc, ctx):
+        ll_yes, _ = rf.loglikelihood(ctx, " yes")
+        ll_no, _ = rf.loglikelihood(ctx, " no")
+        return ll_yes, ll_no
+
+    def process_results(self, doc, results):
+        ll_yes, ll_no = results
+        pred = ll_no > ll_yes
+        gold = doc["label"]
+        return {
+            "acc": pred == gold
+        }
 
-class WNLI(PromptSourceTask):
+    def higher_is_better(self):
+        return {
+            "acc": True
+        }
+
+    def aggregation(self):
+        return {
+            "acc": mean
+        }
+
+
+class WNLI(Task):
     VERSION = 1
     DATASET_PATH = "glue"
     DATASET_NAME = "wnli"
@@ -184,13 +301,49 @@ class WNLI(PromptSourceTask):
         return False
 
     def training_docs(self):
-        return self.dataset["train"]
+        if self._training_docs is None:
+            self._training_docs = list(self.dataset["train"])
+        return self._training_docs
 
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        return "{}\nQuestion: {} True or False?\nAnswer:".format(
+            doc["sentence1"],
+            doc["sentence2"],
+        )
+
+    def doc_to_target(self, doc):
+        # True = entailment
+        # False = not_entailment
+        return " {}".format({0: "False", 1: "True"}[doc["label"]])
+
+    def construct_requests(self, doc, ctx):
+        ll_true, _ = rf.loglikelihood(ctx, " True")
+        ll_false, _ = rf.loglikelihood(ctx, " 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"]
+        return {
+            "acc": pred == gold
+        }
+
+    def higher_is_better(self):
+        return {
+            "acc": True
+        }
+
+    def aggregation(self):
+        return {
+            "acc": mean
+        }
 
-class RTE(PromptSourceTask):
+
+class RTE(Task):
     VERSION = 0
     DATASET_PATH = "glue"
     DATASET_NAME = "rte"
@@ -212,17 +365,45 @@ class RTE(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        return "{}\nQuestion: {} True or False?\nAnswer:".format(
+            doc["sentence1"],
+            doc["sentence2"],
+        )
+
+    def doc_to_target(self, doc):
+        # 0 = entailment
+        # 1 = not_entailment
+        return " {}".format({0: "True", 1: "False"}[doc["label"]])
+
+    def construct_requests(self, doc, ctx):
+        ll_true, _ = rf.loglikelihood(ctx, " True")
+        ll_false, _ = rf.loglikelihood(ctx, " False")
+        return ll_true, ll_false
+
+    def process_results(self, doc, results):
+        ll_true, ll_false = results
+        pred = ll_false > ll_true
+        gold = doc["label"]
+        return {
+            "acc": pred == gold
+        }
+
     def higher_is_better(self):
-        return {"acc": True}
+        return {
+            "acc": True
+        }
 
     def aggregation(self):
-        return {"acc": mean}
+        return {
+            "acc": mean
+        }
 
 
 # Similarity and Paraphrase Tasks
 
 
-class MRPC(PromptSourceTask):
+class MRPC(Task):
     VERSION = 0
     DATASET_PATH = "glue"
     DATASET_NAME = "mrpc"
@@ -244,8 +425,43 @@ class MRPC(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        return "Sentence 1: {}\nSentence 2: {}\nQuestion: Do both sentences mean the same thing?\nAnswer:".format(
+            general_detokenize(doc["sentence1"]),
+            general_detokenize(doc["sentence2"]),
+        )
+
+    def doc_to_target(self, doc):
+        return " {}".format(yesno(doc["label"]))
+
+    def construct_requests(self, doc, ctx):
+        ll_yes, _ = rf.loglikelihood(ctx, " yes")
+        ll_no, _ = rf.loglikelihood(ctx, " no")
+        return ll_yes, ll_no
+
+    def process_results(self, doc, results):
+        ll_yes, ll_no = results
+        gold = doc["label"]
+        pred = ll_yes > ll_no
+        return {
+            "acc": pred == gold,
+            "f1": (gold, pred),
+        }
 
-class QQP(PromptSourceTask):
+    def higher_is_better(self):
+        return {
+            "acc": True,
+            "f1": True
+        }
+
+    def aggregation(self):
+        return {
+            "acc": mean,
+            "f1": f1_score
+        }
+
+
+class QQP(Task):
     VERSION = 0
     DATASET_PATH = "glue"
     DATASET_NAME = "qqp"
@@ -267,6 +483,41 @@ class QQP(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        return "Question 1: {}\nQuestion 2: {}\nQuestion: Do both questions ask the same thing?\nAnswer:".format(
+            doc["question1"],
+            doc["question2"],
+        )
+
+    def doc_to_target(self, doc):
+        return " {}".format(yesno(doc["label"]))
+
+    def construct_requests(self, doc, ctx):
+        ll_yes, _ = rf.loglikelihood(ctx, " yes")
+        ll_no, _ = rf.loglikelihood(ctx, " no")
+        return ll_yes, ll_no
+
+    def process_results(self, doc, results):
+        ll_yes, ll_no = results
+        gold = doc["label"]
+        pred = ll_yes > ll_no
+        return {
+            "acc": pred == gold,
+            "f1": (gold, pred),
+        }
+
+    def higher_is_better(self):
+        return {
+            "acc": True,
+            "f1": True
+        }
+
+    def aggregation(self):
+        return {
+            "acc": mean,
+            "f1": f1_score
+        }
+
 
 class STSB(Task):
     VERSION = 0
@@ -303,7 +554,7 @@ class STSB(Task):
         return " {}".format(doc["label"])
 
     def construct_requests(self, doc, ctx):
-        """Uses RequestFactory to construct Requests and returns an iterable of
+        """ Uses RequestFactory to construct Requests and returns an iterable of 
         Requests which will be sent to the LM.
 
         :param doc:
@@ -314,7 +565,7 @@ class STSB(Task):
             part of the document for `doc`. 
         """
         # TODO: implement evaluation.
-        raise NotImplementedError("Evaluation not implemented")
+        raise NotImplementedError('Evaluation not implemented')
     
     def process_results(self, doc, results):
         """Take a single document and the LM results and evaluates, returning a 
@@ -327,7 +578,7 @@ class STSB(Task):
             The results of the requests created in construct_requests.
         """
         # TODO: implement evaluation.
-        raise NotImplementedError("Evaluation not implemented")
+        raise NotImplementedError('Evaluation not implemented')
 
     def aggregation(self):
         """
@@ -336,7 +587,7 @@ class STSB(Task):
             functions that aggregate a list of metrics
         """
         # TODO: implement evaluation.
-        raise NotImplementedError("Evaluation not implemented")
+        raise NotImplementedError('Evaluation not implemented')
 
     def higher_is_better(self):
         """
@@ -345,4 +596,4 @@ class STSB(Task):
             whether a higher value of the submetric is better
         """
         # TODO: implement evaluation.
-        raise NotImplementedError("Evaluation not implemented")
+        raise NotImplementedError('Evaluation not implemented')

lm_eval/tasks/hans.py

-"""
-Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference
-https://arxiv.org/abs/1902.01007
-
-A controlled evaluation set called HANS (Heuristic Analysis for NLI Systems),
-which contains many examples where the heuristics fail.
-
-Homepage: https://github.com/tommccoy1/hans
-"""
-from lm_eval.base import PromptSourceTask
-
-
-_CITATION = """
-@inproceedings{mccoy-etal-2019-right,
-    title = "Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference",
-    author = "McCoy, Tom  and
-      Pavlick, Ellie  and
-      Linzen, Tal",
-    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
-    month = jul,
-    year = "2019",
-    address = "Florence, Italy",
-    publisher = "Association for Computational Linguistics",
-    url = "https://aclanthology.org/P19-1334",
-    doi = "10.18653/v1/P19-1334",
-    pages = "3428--3448",
-    abstract = "A machine learning system can score well on a given test set by relying on heuristics that are effective for frequent example types but break down in more challenging cases. We study this issue within natural language inference (NLI), the task of determining whether one sentence entails another. We hypothesize that statistical NLI models may adopt three fallible syntactic heuristics: the lexical overlap heuristic, the subsequence heuristic, and the constituent heuristic. To determine whether models have adopted these heuristics, we introduce a controlled evaluation set called HANS (Heuristic Analysis for NLI Systems), which contains many examples where the heuristics fail. We find that models trained on MNLI, including BERT, a state-of-the-art model, perform very poorly on HANS, suggesting that they have indeed adopted these heuristics. We conclude that there is substantial room for improvement in NLI systems, and that the HANS dataset can motivate and measure progress in this area.",
-}
-"""
-
-
-class HANS(PromptSourceTask):
-    VERSION = 0
-    DATASET_PATH = "hans"
-    DATASET_NAME = None
-
-    def has_training_docs(self):
-        return True
-
-    def has_validation_docs(self):
-        return True
-
-    def has_test_docs(self):
-        return False
-
-    def training_docs(self):
-        if self.has_training_docs():
-            # We cache training documents in `self._training_docs` for faster
-            # few-shot processing. If the data is too large to fit in memory,
-            # return the training data as a generator instead of a list.
-            if self._training_docs is None:
-                self._training_docs = list(self.dataset["train"])
-            return self._training_docs
-
-    def validation_docs(self):
-        if self.has_validation_docs():
-            return self.dataset["validation"]
-
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.dataset["test"]

lm_eval/tasks/lama.py

-"""
-https://arxiv.org/abs/1909.01066
-https://arxiv.org/abs/2005.04611
-LAMA is a prob dataset to test the factual and commonsense knowledge in language models. The dataset includes a subset of 
-Google_RE (https://code.google.com/archive/p/relation-extraction-corpus/), TRex (subset of wikidata triples), 
-Conceptnet (https://github.com/commonsense/conceptnet5/wiki) and Squad. 
-
-Homepage: https://github.com/facebookresearch/LAMA
-"""
-from lm_eval.base import PromptSourceTask
-import numpy as np 
-from lm_eval.metrics import mean
-from typing import Optional
-
-_CITATION = """
-@inproceedings{petroni2019language, title={Language Models as Knowledge Bases?},
-               author={F. Petroni, T. Rockt{"{a}}schel, A. H. Miller, P. Lewis, A. Bakhtin, Y. Wu and S. Riedel},
-               booktitle={In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019}, year={2019} }
-
-@inproceedings{petroni2020how,
-               title={How Context Affects Language Models' Factual Predictions},
-               author={Fabio Petroni and Patrick Lewis and Aleksandra Piktus and Tim Rockt{"a}schel and Yuxiang Wu and Alexander H. Miller and Sebastian Riedel},
-               booktitle={Automated Knowledge Base Construction}, year={2020}, url={https://openreview.net/forum?id=025X0zPfn} }
-"""
-
-
-
-class BigScienceLAMA(PromptSourceTask):
-    VERSION = 0
-    DATASET_PATH = "janck/bigscience-lama"
-    DATASET_NAME = None
-
-
-    def has_training_docs(self):
-        # TODO: Fill in the return with `True` if the Task has training data; else `False`.
-        return False
-    def has_validation_docs(self):
-        # TODO: Fill in the return with `True` if the Task has validation data; else `False`.
-        return False
-    def has_test_docs(self):
-        # TODO: Fill in the return with `True` if the Task has test data; else `False`.
-        return True
-    def training_docs(self):
-        if self.has_training_docs():
-            return self.dataset["train"]
-
-
-class Trex(PromptSourceTask):
-    VERSION = 0
-    DATASET_PATH = "lama"
-    DATASET_NAME = "trex"
-
-    def has_training_docs(self):
-        # TODO: Fill in the return with `True` if the Task has training data; else `False`.
-        return False
-
-    def has_validation_docs(self):
-        # TODO: Fill in the return with `True` if the Task has validation data; else `False`.
-        return False
-
-    def has_test_docs(self):
-        # TODO: Fill in the return with `True` if the Task has test data; else `False`.
-        return True
-
-    def training_docs(self):
-        if self.has_training_docs():
-            if self._training_docs is None:
-                self._training_docs = list(self.dataset["train"])
-            return self._training_docs
-
-    def validation_docs(self):
-        if self.has_validation_docs():
-            return self.dataset["validation"]
-
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.dataset["train"]
-
-    def process_results(self, doc, results):
-        out = {}
-        #gold = doc
-        pred = results[0].strip()
-        target = self.doc_to_target(doc)['obj_label']
-        #pred = np.argmax(results)
-        out["acc"] = pred == target
-
-
-        if self.save_examples:
-            example = {
-                "pred": pred,
-                "target": target,
-            }
-            return out, example
-
-        return out
-
-    def higher_is_better(self):
-        return {"acc": True}
-
-    def aggregation(self):
-        return {"acc": mean}
-
-    def doc_to_target(self, doc):
-        return doc
-
-
-class google_re(PromptSourceTask):
-    VERSION = 0
-    DATASET_PATH = "lama"
-    DATASET_NAME = "google_re"
-
-    def has_training_docs(self):
-        # TODO: Fill in the return with `True` if the Task has training data; else `False`.
-        return False
-
-    def has_validation_docs(self):
-        # TODO: Fill in the return with `True` if the Task has validation data; else `False`.
-        return False
-
-    def has_test_docs(self):
-        # TODO: Fill in the return with `True` if the Task has test data; else `False`.
-        return True
-
-    def training_docs(self):
-        if self.has_training_docs():
-            if self._training_docs is None:
-                self._training_docs = list(self.dataset["train"])
-            return self._training_docs
-
-    def validation_docs(self):
-        if self.has_validation_docs():
-            return self.dataset["validation"]
-
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.dataset["train"]
-
-    def process_results(self, doc, results):
-        out = {}
-        pred = results[0].strip()
-
-        target = self.doc_to_target(doc)['obj_label']
-        out["acc"] = pred == target
-
-
-        if self.save_examples:
-            example = {
-                "pred": pred,
-                "target": target,
-            }
-            return out, example
-
-        return out
-
-    def higher_is_better(self):
-        return {"acc": True}
-
-    def aggregation(self):
-        return {"acc": mean}
-
-    def doc_to_target(self, doc):
-        return doc
-
-class Conceptnet(PromptSourceTask):
-    VERSION = 0
-    DATASET_PATH = "lama"
-    DATASET_NAME = "conceptnet"
-
-    def has_training_docs(self):
-        # TODO: Fill in the return with `True` if the Task has training data; else `False`.
-        return False
-
-    def has_validation_docs(self):
-        # TODO: Fill in the return with `True` if the Task has validation data; else `False`.
-        return False
-
-    def has_test_docs(self):
-        # TODO: Fill in the return with `True` if the Task has test data; else `False`.
-        return True
-
-
-    def training_docs(self):
-        if self.has_training_docs():
-            if self._training_docs is None:
-                self._training_docs = list(self.dataset["train"])
-            return self._training_docs
-
-    def validation_docs(self):
-        if self.has_validation_docs():
-            return self.dataset["validation"]
-
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.dataset["train"]
-
-    def process_results(self, doc, results):
-        out = {}
-        pred = results[0].strip()
-
-        target = self.doc_to_target(doc)['obj_label']
-        out["acc"] = pred == target
-
-
-        if self.save_examples:
-            example = {
-                "pred": pred,
-                "target": target,
-            }
-            return out, example
-
-        return out
-
-    def higher_is_better(self):
-        return {"acc": True}
-
-    def aggregation(self):
-        return {"acc": mean}
-
-    def doc_to_target(self, doc):
-        return doc
-
-
-class Squad(PromptSourceTask):
-    VERSION = 0
-    DATASET_PATH = "lama"
-    DATASET_NAME = "squad"
-
-    def has_training_docs(self):
-        # TODO: Fill in the return with `True` if the Task has training data; else `False`.
-        return False
-
-    def has_validation_docs(self):
-        # TODO: Fill in the return with `True` if the Task has validation data; else `False`.
-        return False
-
-    def has_test_docs(self):
-        # TODO: Fill in the return with `True` if the Task has test data; else `False`.
-        return True
-
-
-    def training_docs(self):
-        if self.has_training_docs():
-            if self._training_docs is None:
-                self._training_docs = list(self.dataset["train"])
-            return self._training_docs
-
-    def validation_docs(self):
-        if self.has_validation_docs():
-            return self.dataset["validation"]
-
-    def test_docs(self):
-        if self.has_test_docs():
-
-            self._test_docs = list(self.dataset["train"])
-            return self._test_docs
-
-    def process_results(self, doc, results):
-        out = {}
-        pred = results[0].strip()
-        target = self.doc_to_target(doc)['obj_label']
-        #pred = np.argmax(results)
-        out["acc"] = pred == target
-
-
-        
-        if self.save_examples:
-            example = {
-                "pred": pred,
-                "target": target,
-            }
-            return out, example
-
-        return out
-
-    def higher_is_better(self):
-        return {"acc": True}
-
-    def aggregation(self):
-        return {"acc": mean}
-
-    def doc_to_target(self, doc):
-        return doc
-
-    def max_generation_length(self) -> Optional[int]:
-        """Denote where the max length of the generation if it is obvious from the task."""
-        return 5
-
-

lm_eval/tasks/race.py

@@ -12,7 +12,7 @@ Homepage: https://www.cs.cmu.edu/~glai1/data/race/
 import collections
 import datasets
 import numpy as np
-from lm_eval.base import PromptSourceTask, rf
+from lm_eval.base import rf, Task
 from lm_eval.metrics import mean
 
 
@@ -34,13 +34,13 @@ class each:
         return list(map(self.f, other))
 
 
-class RACE(PromptSourceTask):
+class RACE(Task):
     VERSION = 1
     DATASET_PATH = "race"
     DATASET_NAME = "high"
 
     cache = {}
-    letter_to_num = {"A": 0, "B": 1, "C": 2, "D": 3}
+    letter_to_num = {'A': 0, 'B': 1, 'C': 2, 'D': 3}
 
     def has_training_docs(self):
         return True
@@ -51,88 +51,79 @@ class RACE(PromptSourceTask):
     def has_test_docs(self):
         return True
 
-    # def _collate_data(self, set):
-    #     if set in self.cache:
-    #         return self.cache[set]
-    #     # One big issue with HF's implementation of this dataset: it makes a
-    #     # separate document for each question; meanwhile, in the GPT3 paper it
-    #     # is shown that one document is made per passage.
-
-    #     r = collections.defaultdict(list)
-    #     for item in datasets.load_dataset(
-    #         path=self.DATASET_PATH, name=self.DATASET_NAME
-    #     )[set]:
-    #         r[item["article"]].append(item)
-
-    #     res = list(
-    #         r.values()
-    #         >> each(
-    #             lambda x: {
-    #                 "article": x[0]["article"],
-    #                 "problems": x
-    #                 >> each(
-    #                     lambda y: {
-    #                         "question": y["question"],
-    #                         "answer": y["answer"],
-    #                         "options": y["options"],
-    #                     }
-    #                 ),
-    #             }
-    #         )
-    #     )
-
-    #     self.cache[set] = res
-    #     return res
+    def _collate_data(self, set):
+        if set in self.cache:
+            return self.cache[set]
+        # One big issue with HF's implementation of this dataset: it makes a
+        # separate document for each question; meanwhile, in the GPT3 paper it
+        # is shown that one document is made per passage.
+
+        r = collections.defaultdict(list)
+        for item in datasets.load_dataset(path=self.DATASET_PATH, name=self.DATASET_NAME)[set]:
+            r[item['article']].append(item)
+        
+        res = list(r.values() >> each(lambda x: {
+            'article': x[0]['article'],
+            'problems': x >> each(lambda y: {
+                'question': y['question'],
+                'answer': y['answer'],
+                'options': y['options'],
+            })
+        }))
+
+        self.cache[set] = res
+        return res
 
     def training_docs(self):
-        return self.dataset["train"]
+        return self._collate_data("train")
 
     def validation_docs(self):
-        return self.dataset["validation"]
+        return self._collate_data("validation")
 
     def test_docs(self):
-        return self.dataset["test"]
+        return self._collate_data("test")
 
     @classmethod
     def get_answer_option(cls, problem):
-        answer = cls.letter_to_num[problem["answer"]]
-        return problem["options"][answer]
+        answer = cls.letter_to_num[problem['answer']]
+        return problem['options'][answer]
 
     @classmethod
     def last_problem(cls, doc):
-        return doc["problems"][-1]
-
-    # def doc_to_text(self, doc):
-    #     text = 'Article: ' + doc['article'] + '\n\n'
-    #     for problem in doc['problems'][:-1]:
-    #         if problem['question'][-6:] == '  _  .':
-    #             text += problem['question'][-5:] + self.get_answer_option(problem) + '\n'
-    #         else:
-    #             question = 'Question: ' + problem['question'] + '\n'
-    #             answer = 'Answer: ' + self.get_answer_option(problem) + '\n'
-    #             text += question + answer
-    #     text += self.last_problem(doc)['question']
-    #     return text
-
-    # def doc_to_target(self, doc):
-    #     return " " + self.get_answer_option(self.last_problem(doc))
-
-    # def construct_requests(self, doc, ctx):
-    #     """Uses RequestFactory to construct Requests and returns an iterable of
-    #     Requests which will be sent to the LM.
-
-    #     :param doc:
-    #         The document as returned from training_docs, validation_docs, or test_docs.
-    #     :param ctx: str
-    #         The context string, generated by fewshot_context. This includes the natural
-    #         language description, as well as the few shot examples, and the question
-    #         part of the document for `doc`.
-    #     """
-    #     problem = self.last_problem(doc)
-    #     ll_choices = [
-    #         rf.loglikelihood(ctx, " " + problem["options"][i])[0] for i in range(4)
-    #     ]
-    #     return ll_choices
+        return doc['problems'][-1]
+
+    def doc_to_text(self, doc):
+        text = 'Article: ' + doc['article'] + '\n\n'
+        for problem in doc['problems'][:-1]:
+            if problem['question'][-6:] == '  _  .':
+                text += problem['question'][-5:] + self.get_answer_option(problem) + '\n'
+            else:
+                question = 'Question: ' + problem['question'] + '\n'
+                answer = 'Answer: ' + self.get_answer_option(problem) + '\n'
+                text += question + answer
+        text += self.last_problem(doc)['question']
+        return text
+
+    def doc_to_target(self, doc):
+        return " " + self.get_answer_option(self.last_problem(doc))
+
+    def construct_requests(self, doc, ctx):
+        """ Uses RequestFactory to construct Requests and returns an iterable of 
+        Requests which will be sent to the LM.
+
+        :param doc:
+            The document as returned from training_docs, validation_docs, or test_docs.
+        :param ctx: str
+            The context string, generated by fewshot_context. This includes the natural 
+            language description, as well as the few shot examples, and the question
+            part of the document for `doc`. 
+        """
+        problem = self.last_problem(doc)
+        ll_choices = [
+            rf.loglikelihood(ctx, " " + problem['options'][i])[0]
+            for i in range(4)
+        ]
+        return ll_choices
 
     def process_results(self, doc, results):
         """Take a single document and the LM results and evaluates, returning a 
@@ -144,11 +135,11 @@ class RACE(PromptSourceTask):
         :param results:
             The results of the requests created in construct_requests.
         """
-        #
-        gold = self.letter_to_num[self.doc_to_target(doc)]
-        # gold = self.letter_to_num[self.last_problem(doc)["answer"]]
+        gold = self.letter_to_num[self.last_problem(doc)['answer']]
         pred = np.argmax(results)
-        return {"acc": int(pred == gold)}
+        return {
+            "acc": int(pred == gold)
+        }
 
     def aggregation(self):
         """
@@ -156,7 +147,9 @@ class RACE(PromptSourceTask):
             A dictionary where keys are the names of submetrics and values are 
             functions that aggregate a list of metrics
         """
-        return {"acc": mean}
+        return {
+            "acc": mean
+        }
 
     def higher_is_better(self):
         """
@@ -164,4 +157,6 @@ class RACE(PromptSourceTask):
             A dictionary where keys are the names of submetrics and values are 
             whether a higher value of the submetric is better
         """
-        return {"acc": True}
+        return {
+            "acc": True
+        }

lm_eval/tasks/superglue.py

@@ -12,7 +12,7 @@ TODO: WSC requires free-form generation.
 import numpy as np
 import sklearn
 import transformers.data.metrics.squad_metrics as squad_metrics
-from lm_eval.base import rf, PromptSourceTask
+from lm_eval.base import rf, Task
 from lm_eval.metrics import mean, acc_all, metric_max_over_ground_truths, yesno
 from lm_eval.utils import general_detokenize
 
@@ -32,7 +32,7 @@ _CITATION = """
 """
 
 
-class BoolQ(PromptSourceTask):
+class BoolQ(Task):
     VERSION = 1
     DATASET_PATH = "super_glue"
     DATASET_NAME = "boolq"
@@ -54,8 +54,41 @@ class BoolQ(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        return f"{doc['passage']}\nQuestion: {doc['question']}?\nAnswer:"
     
-class CommitmentBank(PromptSourceTask):
+    def doc_to_target(self, doc):
+        return " " + yesno(doc['label']) 
+
+    def construct_requests(self, doc, ctx):
+
+        ll_yes, _ = rf.loglikelihood(ctx, ' yes')
+        ll_no, _ = rf.loglikelihood(ctx, ' no')
+
+        return ll_yes, ll_no
+
+    def process_results(self, doc, results):
+        ll_yes, ll_no = results
+        gold = doc["label"]
+
+        acc = 1. if (ll_yes > ll_no) == gold else 0.
+
+        return {
+            "acc": acc
+        }
+    
+    def higher_is_better(self):
+        return {
+            "acc": True
+        }
+    
+    def aggregation(self):
+        return {
+            "acc": mean
+        }
+
+
+class CommitmentBank(Task):
     VERSION = 1
     DATASET_PATH = "super_glue"
     DATASET_NAME = "cb"
@@ -77,15 +110,40 @@ class CommitmentBank(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        return "{}\nQuestion: {}. True, False or Neither?\nAnswer:".format(
+            doc["premise"],
+            doc["hypothesis"],
+        )
+
+    def doc_to_target(self, doc):
+        # True = entailment
+        # False = contradiction
+        # Neither = neutral
+        return " {}".format({0: "True", 1: "False", 2: "Neither"}[doc["label"]])
+
+    def construct_requests(self, doc, ctx):
+        ll_true, _ = rf.loglikelihood(ctx, ' True')
+        ll_false, _ = rf.loglikelihood(ctx, ' False')
+        ll_neither, _ = rf.loglikelihood(ctx, ' Neither')
+
+        return ll_true, ll_false, ll_neither
+
     def process_results(self, doc, results):
         gold = doc["label"]
         pred = np.argmax(results)
-        acc = 1.0 if pred == gold else 0.0
+        acc = 1. if pred == gold else 0.
 
-        return {"acc": acc, "f1": (pred, gold)}
+        return {
+            "acc": acc,
+            "f1": (pred, gold)
+        }
     
     def higher_is_better(self):
-        return {"acc": True, "f1": True}
+        return {
+            "acc": True,
+            "f1": True
+        }
 
     @classmethod
     def cb_multi_fi(cls, items):
@@ -105,7 +163,7 @@ class CommitmentBank(PromptSourceTask):
         }
 
 
-class Copa(PromptSourceTask):
+class Copa(Task):
     VERSION = 0
     DATASET_PATH = "super_glue"
     DATASET_NAME = "copa"
@@ -127,25 +185,53 @@ class Copa(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        # Drop the period
+        connector = {
+            "cause": "because",
+            "effect": "therefore",
+        }[doc["question"]]
+        return doc["premise"].strip()[:-1] + f" {connector}"
+
+    def doc_to_target(self, doc):
+        correct_choice = doc["choice1"] if doc["label"] == 0 else doc["choice2"]
+        # Connect the sentences
+        return " " + self.convert_choice(correct_choice)
+
+    def construct_requests(self, doc, ctx):
+        choice1 = " " + self.convert_choice(doc["choice1"])
+        choice2 = " " + self.convert_choice(doc["choice2"])
+        
+        ll_choice1, _ = rf.loglikelihood(ctx, choice1)
+        ll_choice2, _ = rf.loglikelihood(ctx, choice2)
+
+        return ll_choice1, ll_choice2
+
     def process_results(self, doc, results):
         gold = doc["label"]
         pred = np.argmax(results)
-        acc = 1.0 if pred == gold else 0.0
+        acc = 1. if pred == gold else 0.
 
-        return {"acc": acc}
+        return {
+            "acc": acc
+        }
     
     def higher_is_better(self):
-        return {"acc": True}
+        return {
+            "acc": True
+        }
     
     def aggregation(self):
-        return {"acc": mean}
+        return {
+            "acc": mean
+        }
 
     @staticmethod
     def convert_choice(choice):
         return choice[0].lower() + choice[1:]
 
 
-class MultiRC(PromptSourceTask):
+class MultiRC(Task):
     VERSION = 1
     DATASET_PATH = "super_glue"
     DATASET_NAME = "multirc"
@@ -167,19 +253,45 @@ class MultiRC(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        return f"{doc['paragraph']}\nQuestion: {doc['question']}\nAnswer:"
+
+    def doc_to_target(self, doc):
+        return " " + self.format_answer(answer=doc["answer"], label=doc["label"])
+
+    @staticmethod
+    def format_answer(answer, label):
+        label_str = "yes" if label else "no"
+        return f"{answer}\nIs the answer correct? {label_str}"
+
+    def construct_requests(self, doc, ctx):
+        true_choice = self.format_answer(answer=doc["answer"], label=True)
+        false_choice = self.format_answer(answer=doc["answer"], label=False)
+        
+        ll_true_choice, _ = rf.loglikelihood(ctx, f' {true_choice}')
+        ll_false_choice, _ = rf.loglikelihood(ctx, f' {false_choice}')
+
+        return ll_true_choice, ll_false_choice
+
     def process_results(self, doc, results):
         ll_true_choice, ll_false_choice = results
         pred = ll_true_choice > ll_false_choice
-        return {"acc": (pred, doc)}
+        return {
+            "acc": (pred, doc)
+        }
     
     def higher_is_better(self):
-        return {"acc": True}
+        return {
+            "acc": True
+        }
     
     def aggregation(self):
-        return {"acc": acc_all}
+        return {
+            "acc": acc_all
+        }
 
 
-class ReCoRD(PromptSourceTask):
+class ReCoRD(Task):
     VERSION = 0
     DATASET_PATH = "super_glue"
     DATASET_NAME = "record"
@@ -199,31 +311,56 @@ class ReCoRD(PromptSourceTask):
         if self._training_docs is None:
             self._training_docs = []
             for doc in self.dataset["train"]:
-                self._training_docs.append(doc)
+                self._training_docs.append(self._process_doc(doc))
         return self._training_docs
 
     def validation_docs(self):
         # See: training_docs
         for doc in self.dataset["validation"]:
-            yield doc
+            yield self._process_doc(doc)
+
+    @classmethod
+    def _process_doc(cls, doc):
+        return {
+            "passage": doc["passage"],
+            "query": doc["query"],
+            "entities": sorted(list(set(doc["entities"]))),
+            "answers": sorted(list(set(doc["answers"]))),
+        }
+
+    def doc_to_text(self, doc):
+        initial_text, *highlights = doc["passage"].strip().split("\n@highlight\n")
+        text = initial_text + "\n\n"
+        for highlight in highlights:
+            text += f"  - {highlight}.\n"
+        return text
+
+    @classmethod
+    def format_answer(cls, query, entity):
+        return f'  - {query}'.replace("@placeholder", entity)
+
+    def doc_to_target(self, doc):
+        # We only output the first correct entity in a doc
+        return self.format_answer(query=doc["query"], entity=doc["answers"][0])
+
+    def construct_requests(self, doc, ctx):
+        requests = [
+            rf.loglikelihood(ctx, self.format_answer(query=doc["query"], entity=entity))
+            for entity in doc["entities"]
+        ]
+        return requests
 
     def process_results(self, doc, results):
         # ReCoRD's evaluation is actually deceptively simple:
         # - Pick the maximum likelihood prediction entity
         # - Evaluate the accuracy and token F1 PER EXAMPLE
         # - Average over all examples
-
-        # TODO (jon-tow): Look at result
         max_idx = np.argmax(np.array([result[0] for result in results]))
 
         prediction = doc["entities"][max_idx]
         gold_label_set = doc["answers"]
-        f1 = metric_max_over_ground_truths(
-            squad_metrics.compute_f1, prediction, gold_label_set
-        )
-        em = metric_max_over_ground_truths(
-            squad_metrics.compute_exact, prediction, gold_label_set
-        )
+        f1 = metric_max_over_ground_truths(squad_metrics.compute_f1, prediction, gold_label_set)
+        em = metric_max_over_ground_truths(squad_metrics.compute_exact, prediction, gold_label_set)
 
         return {
             "f1": f1,
@@ -243,7 +380,7 @@ class ReCoRD(PromptSourceTask):
         }
 
 
-class WordsInContext(PromptSourceTask):
+class WordsInContext(Task):
     VERSION = 0
     DATASET_PATH = "super_glue"
     DATASET_NAME = "wic"
@@ -265,19 +402,50 @@ class WordsInContext(PromptSourceTask):
     def validation_docs(self):
         return self.dataset["validation"]
 
+    def doc_to_text(self, doc):
+        return "Sentence 1: {}\nSentence 2: {}\nQuestion: Is the word '{}' used in the same way in the" \
+               " two sentences above?\nAnswer:".format(
+                    doc["sentence1"],
+                    doc["sentence2"],
+                    doc["sentence1"][doc["start1"]:doc["end1"]],
+                )
+
+    def doc_to_target(self, doc):
+        return " {}".format({0: "no", 1: "yes"}[doc["label"]])
+
+    def construct_requests(self, doc, ctx):
+        ll_yes, _ = rf.loglikelihood(ctx, ' yes')
+        ll_no, _ = rf.loglikelihood(ctx, ' no')
+
+        return ll_yes, ll_no
+
+    def process_results(self, doc, results):
+        ll_yes, ll_no = results
+        gold = doc["label"]
+
+        acc = 1. if (ll_yes > ll_no) == gold else 0.
+
+        return {
+            "acc": acc
+        }
+
     def higher_is_better(self):
-        return {"acc": True}
+        return {
+            "acc": True
+        }
 
     def aggregation(self):
-        return {"acc": mean}
+        return {
+            "acc": mean
+        }
 
 
-class SGWinogradSchemaChallenge(PromptSourceTask):
+class SGWinogradSchemaChallenge(Task):
     VERSION = 0
     # Note: This implementation differs from Fig G.32 because this is the SuperGLUE,
     #       binary version of the task.
     DATASET_PATH = "super_glue"
-    DATASET_NAME = "wsc.fixed"
+    DATASET_NAME = "wsc"
 
     def has_training_docs(self):
         return True
@@ -293,51 +461,56 @@ class SGWinogradSchemaChallenge(PromptSourceTask):
             if self._training_docs is None:
                 # GPT-3 Paper's format only uses positive examples for fewshot "training"
                 self._training_docs = [
-                    doc for doc in self.dataset["train"] if doc["label"]
+                    doc for doc in
+                    self.dataset["train"]
+                    if doc["label"]
                 ]
             return self._training_docs
 
     def validation_docs(self):
         return self.dataset["validation"]
 
-    def higher_is_better(self):
-        return {"acc": True}
-
-    def aggregation(self):
-        return {"acc": mean}
-
-
-class WinogenderSchemaDiagnostics(PromptSourceTask):
-    VERSION = 0
-    DATASET_PATH = "super_glue"
-    DATASET_NAME = "axg"
-
-    def has_training_docs(self):
-        return False
+    def doc_to_text(self, doc):
+        raw_passage = doc["text"]
+        # NOTE: HuggingFace span indices are word-based not character-based.
+        pre = " ".join(raw_passage.split()[:doc["span2_index"]])
+        post = raw_passage[len(pre) + len(doc["span2_text"]) + 1:]
+        passage = general_detokenize(pre + " *{}*".format(doc['span2_text']) + post)
+        noun = doc["span1_text"]
+        pronoun = doc["span2_text"]
+        text = (
+            f"Passage: {passage}\n"
+            + f"Question: In the passage above, does the pronoun \"*{pronoun}*\" refer to \"*{noun}*\"?\n"
+            + "Answer:"
+        )
+        return text
 
-    def has_validation_docs(self):
-        return False 
+    def doc_to_target(self, doc):
+        return " " + yesno(doc['label'])
 
-    def has_test_docs(self):
-        return True 
+    def construct_requests(self, doc, ctx):
 
-    def test_docs(self):
-        return self.dataset["test"]
+        ll_yes, _ = rf.loglikelihood(ctx, ' yes')
+        ll_no, _ = rf.loglikelihood(ctx, ' no')
 
+        return ll_yes, ll_no
 
-class BroadcoverageDiagnostics(PromptSourceTask):
-    VERSION = 0
-    DATASET_PATH = "super_glue"
-    DATASET_NAME = "axb"
+    def process_results(self, doc, results):
+        ll_yes, ll_no = results
+        gold = doc["label"]
 
-    def has_training_docs(self):
-        return False
+        acc = 1. if (ll_yes > ll_no) == gold else 0.
 
-    def has_validation_docs(self):
-        return False 
+        return {
+            "acc": acc
+        }
 
-    def has_test_docs(self):
-        return True 
+    def higher_is_better(self):
+        return {
+            "acc": True
+        }
 
-    def test_docs(self):
-        return self.dataset["test"]
+    def aggregation(self):
+        return {
+            "acc": mean
+        }

lm_eval/tasks/wino_bias.py

-"""
-Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods
-https://arxiv.org/abs/1804.06876
-
-Winograd-schema evaluation of gendered coreference resolution.
-The dataset contains pro-stereotypical and anti-stereotypical parts. The difference in accuracy for those two subsets
-quatnifies bias.
-
-Homepage: https://uclanlp.github.io/corefBias/overview
-"""
-from lm_eval.base import PromptSourceTask, mean
-import transformers.data.metrics.squad_metrics as squad_metrics
-
-_CITATION = """
-@inproceedings{zhao-etal-2018-gender,
-    title = "Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods",
-    author = "Zhao, Jieyu  and
-      Wang, Tianlu  and
-      Yatskar, Mark  and
-      Ordonez, Vicente  and
-      Chang, Kai-Wei",
-    booktitle = "Proceedings of the 2018 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers)",
-    month = jun,
-    year = "2018",
-    address = "New Orleans, Louisiana",
-    publisher = "Association for Computational Linguistics",
-    url = "https://aclanthology.org/N18-2003",
-    doi = "10.18653/v1/N18-2003",
-    pages = "15--20",
-    abstract = "In this paper, we introduce a new benchmark for co-reference resolution focused on gender bias, WinoBias. Our corpus contains Winograd-schema style sentences with entities corresponding to people referred by their occupation (e.g. the nurse, the doctor, the carpenter). We demonstrate that a rule-based, a feature-rich, and a neural coreference system all link gendered pronouns to pro-stereotypical entities with higher accuracy than anti-stereotypical entities, by an average difference of 21.1 in F1 score. Finally, we demonstrate a data-augmentation approach that, in combination with existing word-embedding debiasing techniques, removes the bias demonstrated by these systems in WinoBias without significantly affecting their performance on existing datasets.",
-}
-"""
-
-
-class WinoBias(PromptSourceTask):
-    VERSION = 0
-    DATASET_PATH = "wino_bias"
-
-    def has_training_docs(self):
-        return False
-
-    def has_validation_docs(self):
-        return True
-
-    def has_test_docs(self):
-        return True
-
-    def training_docs(self):
-        pass
-
-    def validation_docs(self):
-        return self.dataset["validation"]
-
-    def test_docs(self):
-        return self.dataset["test"]
-
-    def process_results(self, doc, results):
-        """Take a single document and the LM results and evaluates, returning a
-        dict where keys are the names of submetrics and values are the values of
-        the metric for that one document
-
-        :param doc:
-            The document as returned from training_docs, validation_docs, or test_docs.
-        :param results:
-            The results of the requests created in construct_requests.
-        """
-        target = self.doc_to_target(doc).strip()
-        pred = " ".join(results[0].strip().split(" ")[:len(target.split(" "))])
-
-        # The original paper uses F1. In the case of exactly one predicted and one gold mention,
-        # F1 and exact match are equivalent.
-        em = squad_metrics.compute_exact(target, pred)
-        out = {"em": em}
-
-        if self.save_examples:
-            example = {"target": target, "pred": pred}
-            return out, example
-        return out
-
-    def aggregation(self):
-        """
-        :returns: {str: [metric_score] -> float}
-            A dictionary where keys are the names of submetrics and values are
-            functions that aggregate a list of metric scores
-        """
-        return {'em': mean}
-
-    def higher_is_better(self):
-        return {'em': True}
-
-
-class WinoBiasType1Pro(WinoBias):
-    DATASET_NAME = "type1_pro"
-
-
-class WinoBiasType1Anti(WinoBias):
-    DATASET_NAME = "type1_anti"
-
-
-class WinoBiasType2Pro(WinoBias):
-    DATASET_NAME = "type2_pro"
-
-
-class WinoBiasType2Anti(WinoBias):
-    DATASET_NAME = "type2_anti"

lm_eval/utils.py

@@ -146,19 +146,6 @@ class Reorderer:
         
         return res
 
-
-def flatten(d, parent_key='', sep='_'):
-    # From: https://stackoverflow.com/a/6027615
-    items = []
-    for k, v in d.items():
-        new_key = parent_key + sep + k if parent_key else k
-        if isinstance(v, collections.MutableMapping):
-            items.extend(flatten(v, new_key, sep=sep).items())
-        else:
-            items.append((new_key, v))
-    return dict(items)
-
-
 def positional_deprecated(fn):
     """
     A decorator to nudge users into passing only keyword args (`kwargs`) to the 

main.py

@@ -9,18 +9,18 @@ logging.getLogger("openai").setLevel(logging.WARNING)
 
 def parse_args():
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model", required=True)
-    parser.add_argument("--model_args", default="")
-    parser.add_argument("--tasks", default="all_tasks")
-    parser.add_argument("--provide_description", action="store_true")
-    parser.add_argument("--num_fewshot", type=int, default=0)
-    parser.add_argument("--batch_size", type=int, default=None)
-    parser.add_argument("--device", type=str, default=None)
-    parser.add_argument("--output_path", default=None)
-    parser.add_argument("--limit", type=int, default=None)
-    parser.add_argument("--no_cache", action="store_true")
-    parser.add_argument("--description_dict_path", default=None)
-    parser.add_argument("--check_integrity", action="store_true")
+    parser.add_argument('--model', required=True)
+    parser.add_argument('--model_args', default="")
+    parser.add_argument('--tasks', default="all_tasks")
+    parser.add_argument('--provide_description', action="store_true")
+    parser.add_argument('--num_fewshot', type=int, default=0)
+    parser.add_argument('--batch_size', type=int, default=None)
+    parser.add_argument('--device', type=str, default=None)
+    parser.add_argument('--output_path', default=None)
+    parser.add_argument('--limit', type=int, default=None)
+    parser.add_argument('--no_cache', action="store_true")
+    parser.add_argument('--description_dict_path', default=None)
+    parser.add_argument('--check_integrity', action="store_true")
     return parser.parse_args()
 
 
@@ -29,9 +29,7 @@ def main():
     assert not args.provide_description  # not implemented
     
     if args.limit:
-        print(
-            "WARNING: --limit SHOULD ONLY BE USED FOR TESTING. REAL METRICS SHOULD NOT BE COMPUTED USING LIMIT."
-        )
+        print("WARNING: --limit SHOULD ONLY BE USED FOR TESTING. REAL METRICS SHOULD NOT BE COMPUTED USING LIMIT.")
 
     if args.tasks == "all_tasks":
         task_names = tasks.ALL_TASKS
@@ -40,7 +38,7 @@ def main():
 
     description_dict = {}
     if args.description_dict_path:
-        with open(args.description_dict_path, "r") as f:
+        with open(args.description_dict_path, 'r') as f:
             description_dict = json.load(f)
 
     results = evaluator.simple_evaluate(
@@ -53,10 +51,9 @@ def main():
         no_cache=args.no_cache,
         limit=args.limit,
         description_dict=description_dict,
-        check_integrity=args.check_integrity,
+        check_integrity=args.check_integrity
     )
 
-    print(results)
     dumped = json.dumps(results, indent=2)
     
     print(dumped)