Merge branch 'master' of github.com:EleutherAI/lm_evaluation_harness into cfsquad

# Conflicts:
#	lm_eval/tasks/squad.py

lm_eval/tasks/pubmedqa.py

 import numpy as np
-import json
-import random
-from .common import HFTask 
+from .common import HFTask
 from lm_eval.base import rf
 from ..metrics import mean
 
@@ -40,12 +38,6 @@ class Pubmed_QA(HFTask):
     def doc_to_target(self, doc):
         return " {}".format(doc["final_decision"])
 
-    def fewshot_examples(self, k):
-        # Since only test docs sample from test docs
-        if self._training_docs is None:
-            self._training_docs = list(self.test_docs())
-        return random.sample(self._training_docs, k)
-
     def construct_requests(self, doc, ctx):
         """ Uses RequestFactory to construct Requests and returns
         an iterable of Requests which will be sent to the LM.

lm_eval/tasks/qa4mre.py

 import os
-import numpy as np
-from best_download import download_file
-from lm_eval.base import MultipleChoiceTask, rf
-from lm_eval.metrics import mean
 import xml.etree.ElementTree as ET
-import random
+from best_download import download_file
+from lm_eval.base import MultipleChoiceTask
+
 
 class QA4MRE(MultipleChoiceTask):
     YEAR = None
@@ -46,12 +44,6 @@ class QA4MRE(MultipleChoiceTask):
     def has_test_docs(self):
         return True
 
-    def fewshot_examples(self, k):
-        # Since only test docs sample from test docs
-        if self._training_docs is None:
-            self._training_docs = list(self.test_docs())
-        return random.sample(self._training_docs, k)
-
     def _convert_standard(self, question):
         choices = [i.text for i in question.iter('answer')]
         out_doc = {

lm_eval/tasks/quac.py

 import json
-import random
 import os
 from lm_eval.base import Task
 from ..utils import sh

lm_eval/tasks/race.py

@@ -5,11 +5,6 @@ from lm_eval.base import rf
 from ..metrics import mean
 from . common import HFTask
 
-import os
-from functools import reduce
-import operator
-from tqdm import tqdm
-import json
 
 class each:
     def __init__(self, f):

lm_eval/tasks/sat.py

-import json
-import random
 import os
-from lm_eval.base import MultipleChoiceTask, rf
-from ..metrics import mean
-from tqdm import auto as tqdm_lib
-from . common import simple_accuracy_metric
-import numpy as np
-from ..utils import sh
+from lm_eval.base import MultipleChoiceTask
 
 
 class SATAnalogies(MultipleChoiceTask):    

lm_eval/tasks/sciq.py

 import os
 import json
-from ..utils import sh
-from lm_eval.base import MultipleChoiceTask, rf
-from ..metrics import mean
 import zipfile
+from lm_eval.base import MultipleChoiceTask
 from best_download import download_file
 
 

lm_eval/tasks/squad.py

@@ -17,7 +17,6 @@ def _squad_agg(key, items):
     return _squad_metric(predictions=predictions, references=references)[key]
 
 
-
 class SQuAD2(HFTask):
     DATASET_PATH = "squad_v2"
     DATASET_NAME = None

lm_eval/tasks/storycloze.py

-import json
-import random
-from lm_eval.base import Task
-from ..utils import sh
 import csv
+from lm_eval.base import Task
+
 
 class StoryCloze(Task):
     NEEDS_MANUAL_DL = True

lm_eval/tasks/superglue.py

@@ -4,11 +4,11 @@ To-do:
     - ReCoRD
 """
 import numpy as np
+import sklearn
+import transformers.data.metrics.squad_metrics as squad_metrics
 from . common import HFTask, yesno
 from lm_eval.base import rf
 from ..metrics import mean, acc_all, metric_max_over_ground_truths
-import sklearn
-import transformers.data.metrics.squad_metrics as squad_metrics
 from ..utils import general_detokenize
 
 
@@ -23,7 +23,7 @@ class BoolQ(HFTask):
         return True
 
     def has_test_docs(self):
-        return True
+        return False
 
     def fewshot_description(self):
         # TODO: figure out actual description
@@ -74,7 +74,7 @@ class CommitmentBank(HFTask):
         return True
 
     def has_test_docs(self):
-        return True
+        return False
 
     def fewshot_description(self):
         # TODO: figure out actual description
@@ -145,7 +145,7 @@ class Copa(HFTask):
         return True
 
     def has_test_docs(self):
-        return True
+        return False
 
     def fewshot_description(self):
         # TODO: figure out actual description
@@ -209,7 +209,7 @@ class MultiRC(HFTask):
         return True
 
     def has_test_docs(self):
-        return True
+        return False
 
     def fewshot_description(self):
         # TODO: figure out actual description
@@ -355,7 +355,7 @@ class WordsInContext(HFTask):
         return True
 
     def has_test_docs(self):
-        return True
+        return False
 
     def fewshot_description(self):
         # TODO: figure out actual description
@@ -412,7 +412,7 @@ class SGWinogradSchemaChallenge(HFTask):
         return True
 
     def has_test_docs(self):
-        return True
+        return False
 
     def training_docs(self):
         if self.has_training_docs():

lm_eval/tasks/translation.py

-import abc
-import json
-import random
-import os
-from collections import Iterable
-from pprint import pprint
-
 import pycountry
+from pprint import pprint
 from sacrebleu import sacrebleu
-import logging
-
 from lm_eval import metrics
 from lm_eval.base import Task, rf
 
@@ -86,11 +78,14 @@ class GeneralTranslationTask(Task):
         } for src, ref in zip(self.src_data, self.ref_data)]
 
     def doc_to_text(self, doc):
-        return doc["src"]
+        language_codes = self.sacrebleu_language_pair.split("-")
+        src_lang = code_to_language(language_codes[0])
+        tar_lang = code_to_language(language_codes[1])
+        return f"{src_lang} phrase: " + doc["src"] + f"\n{tar_lang} phrase:"
 
     def doc_to_target(self, doc):
         # This shows a single target, though there may be multiple targets in a lang test
-        return doc["ref"] if isinstance(doc["ref"], str) else doc["ref"][0]
+        return " " + doc["ref"] if isinstance(doc["ref"], str) else doc["ref"][0]
 
     def construct_requests(self, doc, ctx):
         """ Uses RequestFactory to construct Requests and returns an iterable of

lm_eval/tasks/triviaqa.py

 import os
 import json
-import random
 from lm_eval.base import Task, rf
 from ..metrics import mean
 from ..utils import sh
 
+
 class TriviaQA(Task):
     def download(self):
         if not os.path.exists('data/triviaqa'):

lm_eval/tasks/unscramble.py

 import gzip
 import json
-import random
 import shutil
 from pathlib import Path
 from best_download import download_file

lm_eval/tasks/wikitext.py

-import numpy as np
-from scipy.stats import pearsonr, spearmanr
-from sklearn.metrics import f1_score, matthews_corrcoef
-from tqdm import auto as tqdm_lib
-from . common import NLP_TASK, simple_accuracy_metric, yesno
+from . common import HFTask
 
-class WikiText103(NLP_TASK):
+
+class WikiText103(HFTask):
     NLP_PATH = "wikitext"
     NLP_NAME = "wikitext-103-raw-v1"
 
@@ -66,7 +63,7 @@ class WikiText103(NLP_TASK):
         raise NotImplementedError('Evaluation not implemented')
 
 
-class WikiText2(NLP_TASK):
+class WikiText2(HFTask):
     NLP_PATH = "wikitext"
     NLP_NAME = "wikitext-2-raw-v1"
 

lm_eval/tasks/wsc273.py

@@ -56,10 +56,14 @@ class WinogradSchemaChallenge273(HFTask):
         # TODO: redo description
         return "Winograd schema sentence with correct continuation. True. Winograd schema sentence with incorrect continuation. False."
 
-    def fewshot_examples(self, k):
+    def fewshot_examples(self, k, rnd):
         # NOTE: `super().fewshot_examples` samples from training docs which are
         # not available for this test-set-only dataset.
-        return random.sample(list(self.test_docs()), k)
+
+        if self._fewshot_docs is None:
+            self._fewshot_docs = list(self.test_docs())
+
+        return rnd.sample(list(self._fewshot_docs), k)
 
     def doc_to_text(self, doc):
         return self.partial_context(doc, doc["options"][doc["label"]])

lm_eval/utils.py

 import os
 import re
+import collections
 
 
 class ExitCodeError(Exception):
@@ -42,6 +43,14 @@ def chunks(iter, n):
     
     if arr: yield arr
 
+def group(arr, fn):
+    res = collections.defaultdict(list)
+
+    for ob in arr:
+        res[fn(ob)].append(ob)
+    
+    return list(res.values())
+
 def general_detokenize(string):
     string = string.replace(" n't", "n't")
     string = string.replace(" )", ")")
@@ -49,4 +58,34 @@ def general_detokenize(string):
     string = string.replace("\" ", "\"")
     string = string.replace(" \"", "\"")
     string = re.sub(r" (['.,])", r"\1", string)
-    return string
\ No newline at end of file
+    return string
+
+
+class Reorderer:
+    def __init__(self, arr, fn):
+        self.size = len(arr)
+        arr = list(enumerate(arr))
+        arr = group(arr, lambda x: fn(x[1]))
+        arr = [
+            ([y[0] for y in x], x[0][1]) for x in arr
+        ]
+        arr.sort(key=lambda x: fn(x[1]))
+
+        self.arr = arr
+        
+    
+    def get_reordered(self):
+        return [x[1] for x in self.arr]
+    
+    def get_original(self, newarr):
+        res = [None] * self.size
+        cov = [False] * self.size
+
+        for (inds, _), v in zip(self.arr, newarr):
+            for ind in inds: 
+                res[ind] = v
+                cov[ind] = True
+        
+        assert all(cov)
+        
+        return res
\ No newline at end of file

main.py

@@ -2,8 +2,6 @@ import argparse
 import json
 import numpy as np
 import random
-import itertools
-import collections
 import logging
 
 from lm_eval import models, tasks, evaluator, base
@@ -35,7 +33,7 @@ def main():
         print("WARNING: --limit SHOULD ONLY BE USED FOR TESTING. REAL METRICS SHOULD NOT BE COMPUTED USING LIMIT.")
 
     if not args.no_cache:
-        lm = base.CachingLM(lm, 'lm_cache/' + args.model + '_' + args.model_args.replace('=', '-').replace(',', '_') + '.db')
+        lm = base.CachingLM(lm, 'lm_cache/' + args.model + '_' + args.model_args.replace('=', '-').replace(',', '_').replace('/', '-') + '.db')
     if args.tasks == "all_tasks":
         task_names = tasks.ALL_TASKS
     else:

scripts/cost_estimate.py

-import argparse
-import json
-import numpy as np
 import random
-import itertools
-import collections
-import logging
-
-from lm_eval import models, tasks, evaluator, base
-import random
-from lm_eval.base import LM
 import transformers
+from lm_eval import tasks, evaluator
+from lm_eval.base import LM
 
 
 class DryrunLM(LM):

scripts/fewshot_description_experiment.py

-import argparse
 import json
 import numpy as np
 import random
-import itertools
-import collections
 import logging
-
 from lm_eval import models, tasks, evaluator, base
 
 logging.getLogger("openai").setLevel(logging.WARNING)

scripts/write_out.py

@@ -2,7 +2,6 @@ import argparse
 import numpy as np
 import os
 import random
-
 from lm_eval import tasks
 from lm_eval.utils import join_iters
 
@@ -16,14 +15,13 @@ def parse_args():
     parser.add_argument('--provide_description', action="store_true")
     parser.add_argument('--sets', type=str, default="val") # example: val,test
     parser.add_argument('--num_fewshot', type=int, default=1)
-    parser.add_argument('--seed', type=int, default=1234)
+    parser.add_argument('--seed', type=int, default=42)
     parser.add_argument('--num_examples', type=int, default=1)
     return parser.parse_args()
 
 
 def main():
     args = parse_args()
-    random.seed(args.seed)
     np.random.seed(args.seed)
 
     if args.tasks == "all_tasks":
@@ -33,6 +31,8 @@ def main():
     task_dict = tasks.get_task_dict(task_names)
     os.makedirs(args.output_base_path, exist_ok=True)
     for task_name, task in task_dict.items():
+        rnd = random.Random()
+        rnd.seed(args.seed)
 
         iters = []
 
@@ -54,6 +54,7 @@ def main():
                     doc=doc,
                     provide_description=args.provide_description,
                     num_fewshot=args.num_fewshot,
+                    rnd=rnd
                 )
                 f.write(ctx + "\n")
 

tests/test_models.py

 import lm_eval.models as models
-import lm_eval.base as base
+
 
 def test_gpt2():
     gpt2 = models.get_model('gpt2').create_from_arg_string("device=cpu")