Merge branch 'master' into thomas/fix_best_download_version

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"
 
@@ -26,12 +26,8 @@ class BoolQ(HFTask):
     def has_test_docs(self):
         return False
 
-    def fewshot_description(self):
-        # TODO: figure out actual description
-        return "Read the following passages and answer each question with a yes or a no."
-
     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']) 
@@ -78,11 +74,6 @@ class CommitmentBank(HFTask):
     def has_test_docs(self):
         return False
 
-    def fewshot_description(self):
-        # TODO: figure out actual description
-        return "Given a premise and a hypothesis, classify whether the author of the premise is committed" \
-            "to the truth of the hypothesis. The three possible labels are true, false or neither."
-
     def doc_to_text(self, doc):
         return "{}\nQuestion: {}. True, False or Neither?\nAnswer:".format(
             doc["premise"],
@@ -150,11 +141,6 @@ class Copa(HFTask):
     def has_test_docs(self):
         return False
 
-    def fewshot_description(self):
-        # TODO: figure out actual description
-        return "Given a premise and one alternative with a causal relation to the premise and another without," \
-            "choose the more plausible alternative"
-
     def doc_to_text(self, doc):
         # Drop the period
         connector = {
@@ -202,7 +188,7 @@ class Copa(HFTask):
 
 
 class MultiRC(HFTask):
-    VERSION = 0
+    VERSION = 1
     DATASET_PATH = "super_glue"
     DATASET_NAME = "multirc"
 
@@ -215,10 +201,6 @@ class MultiRC(HFTask):
     def has_test_docs(self):
         return False
 
-    def fewshot_description(self):
-        # TODO: figure out actual description
-        return "READING COMPREHENSION ANSWER KEY"
-
     def doc_to_text(self, doc):
         return f"{doc['paragraph']}\nQuestion: {doc['question']}\nAnswer:"
 
@@ -228,7 +210,7 @@ class MultiRC(HFTask):
     @staticmethod
     def format_answer(answer, label):
         label_str = "yes" if label else "no"
-        return f"{label_str}, {answer}"
+        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)
@@ -240,7 +222,8 @@ class MultiRC(HFTask):
         return ll_true_choice, ll_false_choice
 
     def process_results(self, doc, results):
-        pred = np.argmax(results)
+        ll_true_choice, ll_false_choice = results
+        pred = ll_true_choice > ll_false_choice
         return {
             "acc": (pred, doc)
         }
@@ -270,10 +253,6 @@ class ReCoRD(HFTask):
     def has_test_docs(self):
         return False
 
-    def fewshot_description(self):
-        # TODO: figure out actual description
-        return ""
-
     def training_docs(self):
         # In ReCoRD, each doc manifests multiple "examples" in the context of few shot example packing.
         # Each doc consists of multiple answer candidates, each of which is scored yes/no.
@@ -363,10 +342,6 @@ class WordsInContext(HFTask):
     def has_test_docs(self):
         return False
 
-    def fewshot_description(self):
-        # TODO: figure out actual description
-        return ""
-
     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(
@@ -432,12 +407,6 @@ class SGWinogradSchemaChallenge(HFTask):
                 ]
             return self._training_docs
 
-    def fewshot_description(self):
-        return "Final Exam with Answer Key\n" \
-           "Instructions: Please carefully read the following passages. " \
-           "For each passage, you must identify which noun the pronoun marked in *bold*" \
-           " refers to.\n====="
-
     def doc_to_text(self, doc):
         raw_passage = doc["text"]
         # NOTE: HuggingFace span indices are word-based not character-based.

lm_eval/tasks/translation.py

@@ -166,12 +166,6 @@ class GeneralTranslationTask(Task):
             "ter": False,
         }
 
-    def fewshot_description(self):
-        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"Translate these {src_lang} phrases to {tar_lang}."
-
     def __str__(self):
         language_codes = self.sacrebleu_language_pair.split("-")
         src_lang = code_to_language(language_codes[0])

lm_eval/tasks/triviaqa.py

@@ -36,10 +36,6 @@ class TriviaQA(Task):
     def test_docs(self):
         raise NotImplementedError()
     
-    def fewshot_description(self):
-        # TODO: figure out fewshot description
-        return ""
-    
     def doc_to_text(self, doc):
         return f"Question: {doc['Question']}\nAnswer:"
 
@@ -56,7 +52,6 @@ class TriviaQA(Task):
                 ret.append(alias)
 
         return ret
-        
 
     def construct_requests(self, doc, ctx):
         ret = []

lm_eval/tasks/truthfulqa.py

@@ -85,9 +85,14 @@ class TruthfulQAMultipleChoice(Task):
     def doc_to_target(self, doc):
         return " "
 
-    def fewshot_context(self, doc, num_fewshot, provide_description, rnd):
+    def fewshot_context(self, doc, num_fewshot, provide_description=None, rnd=None, description=None):
         assert num_fewshot == 0, "TruthfulQA is intended only for the zero-shot setting."
-        return super().fewshot_context(doc, num_fewshot, provide_description, rnd)
+        return super().fewshot_context(
+            doc=doc,
+            num_fewshot=num_fewshot,
+            rnd=rnd,
+            description=description
+        )
 
     def construct_requests(self, doc, ctx):
         """ Uses RequestFactory to construct Requests and returns an iterable of
@@ -217,9 +222,14 @@ class TruthfulQAGeneration(Task):
     def doc_to_target(self, doc):
         return " "
 
-    def fewshot_context(self, doc, num_fewshot, provide_description, rnd):
+    def fewshot_context(self, doc, num_fewshot, provide_description=None, rnd=None, description=None):
         assert num_fewshot == 0, "TruthfulQA is intended only for the zero-shot setting."
-        return super().fewshot_context(doc, num_fewshot, provide_description, rnd)
+        return super().fewshot_context(
+            doc=doc,
+            num_fewshot=num_fewshot,
+            rnd=rnd,
+            description=description
+        )
 
     def construct_requests(self, doc, ctx):
         """ Uses RequestFactory to construct Requests and returns an iterable of

lm_eval/tasks/unscramble.py

@@ -45,9 +45,6 @@ class WordUnscrambleTask(Task):
         file = self.BASE_PATH / self.FILENAME
         return (json.loads(line) for line in open(file).read().splitlines())
 
-    def fewshot_description(self):
-        return "Please unscramble the letters into a word, and write that word:"
-
     def doc_to_text(self, doc):
         return doc["context"]
 

lm_eval/tasks/webqs.py

@@ -17,10 +17,6 @@ class WebQs(HFTask):
     def has_test_docs(self):
         return True
 
-    def fewshot_description(self):
-        # TODO: figure out description
-        return ""
-
     def doc_to_text(self, doc):
         return "Question: " + doc['question'] + '\nAnswer:'
 
@@ -40,7 +36,6 @@ class WebQs(HFTask):
                 ret.append(alias)
 
         return ret
-        
 
     def construct_requests(self, doc, ctx):
         ret = []
@@ -62,4 +57,4 @@ class WebQs(HFTask):
     def higher_is_better(self):
         return {
             "acc": True
-        }
+        }
\ No newline at end of file

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'):
@@ -49,10 +49,6 @@ class WikiText(PerplexityTask):
             download_file("https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip", local_file="data/wikitext/wikitext-2-raw-v1.zip", expected_checksum="ef7edb566e3e2b2d31b29c1fdb0c89a4cc683597484c3dc2517919c615435a11")
             sh("cd data/wikitext/ && unzip wikitext-2-raw-v1.zip")
 
-    def fewshot_description(self):
-        # TODO: figure out fewshot description
-        return ""
-
     def has_validation_docs(self):
         return True
 
@@ -87,4 +83,4 @@ class WikiText(PerplexityTask):
     
     def count_words(self, doc):
         # count number of words in *original doc before detokenization*
-        return len(re.split(r"\s+", doc))
+        return len(re.split(r"\s+", doc))
\ No newline at end of file

lm_eval/tasks/winogrande.py

@@ -29,10 +29,6 @@ class Winogrande(HFTask):
     def doc_to_text(self, doc):
         return self.partial_context(doc, doc["option" + doc["answer"]])
 
-    def fewshot_description(self):
-        # TODO: redo description
-        return "Winograd schema sentence including a either a ___ blank with a missing word, making the pronoun ambiguous, or the same with the word filled in."
-
     @classmethod
     def partial_context(cls, doc, option):
         # Substitute the pronoun in the sentence with the specified option

lm_eval/tasks/wsc273.py

@@ -53,10 +53,6 @@ class WinogradSchemaChallenge273(HFTask):
     def has_test_docs(self):
         return True
 
-    def fewshot_description(self):
-        # TODO: redo description
-        return "Winograd schema sentence with correct continuation. True. Winograd schema sentence with incorrect continuation. False."
-
     def fewshot_examples(self, k, rnd):
         # NOTE: `super().fewshot_examples` samples from training docs which are
         # not available for this test-set-only dataset.

lm_eval/utils.py

 import os
 import re
 import collections
+import functools
+import inspect
 
 
 class ExitCodeError(Exception):
@@ -138,4 +140,18 @@ class Reorderer:
         
         assert all(cov)
         
-        return res
+        return res
\ No newline at end of file
+
+def positional_deprecated(fn):
+    """
+    A decorator to nudge users into passing only keyword args (`kwargs`) to the 
+    wrapped function, `fn`.
+    """
+    @functools.wraps(fn)
+    def _wrapper(*args, **kwargs):
+        if len(args) != 1 if inspect.ismethod(fn) else 0: 
+            print(f"WARNING: using {fn.__name__} with positional arguments is "
+                "deprecated and will be disallowed in a future version of "
+                "lm-evaluation-harness!")
+        return fn(*args, **kwargs)
+    return _wrapper

main.py

@@ -19,6 +19,7 @@ def parse_args():
     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)
     return parser.parse_args()
 
 
@@ -34,15 +35,21 @@ def main():
     else:
         task_names = args.tasks.split(",")
 
+    description_dict = {}
+    if args.description_dict_path:
+        with open(args.description_dict_path, 'r') as f:
+            description_dict = json.load(f)
+
     results = evaluator.simple_evaluate(
         model=args.model,
         model_args=args.model_args,
-        task_names=task_names,
+        tasks=task_names,
         num_fewshot=args.num_fewshot,
         batch_size=args.batch_size,
         device=args.device,
         no_cache=args.no_cache,
         limit=args.limit,
+        description_dict=description_dict
     )
 
     dumped = json.dumps(results, indent=2)

scripts/cost_estimate.py

@@ -51,7 +51,14 @@ def main():
     values = []
     for taskname in task_list.split(","):
         lm.tokencost = 0
-        evaluator.evaluate(lm, {taskname: tasks.get_task(taskname)()}, False, 0, None, bootstrap_iters=10)
+        evaluator.evaluate(
+            lm=lm,
+            task_dict={taskname: tasks.get_task(taskname)()},
+            num_fewshot=0,
+            limit=None,
+            bootstrap_iters=10,
+            description_dict=None
+        )
 
         print(taskname, lm.tokencost)
         values.append([taskname, lm.tokencost, lm.tokencost / 1000 * 0.0008, lm.tokencost / 1000 * 0.0012, lm.tokencost / 1000 * 0.006, lm.tokencost / 1000 * 0.06])

scripts/fewshot_description_experiment.py

-import json
-import numpy as np
-import random
-import logging
-from lm_eval import models, tasks, evaluator, base
-
-logging.getLogger("openai").setLevel(logging.WARNING)
-
-
-fewshot_descriptions = [
-    "foo",
-    "bar"
-]
-
-task = "lambada"
-num_fewshot = 0
-model = "gpt2"
-model_args = ""
-limit = None
-no_cache = False
-
-
-class CustomDescTask:
-    def __init__(self, task, desc):
-        self.task = task
-        self.desc = desc
-
-        def fewshot_description():
-            return self.desc
-        
-        self.task.fewshot_description = fewshot_description
-
-    def __getattr__(self, attr):
-        return getattr(self.task, attr)
-
-
-def main():
-    random.seed(42)
-    np.random.seed(42)
-
-    lm = models.get_model(model).create_from_arg_string(model_args)
-    
-    if limit:
-        print("WARNING: --limit SHOULD ONLY BE USED FOR TESTING. REAL METRICS SHOULD NOT BE COMPUTED USING LIMIT.")
-
-    if not no_cache:
-        lm = base.CachingLM(lm, 'lm_cache/' + model + '_' + model_args.replace('=', '-').replace(',', '_') + '.db')
-
-    task_dict = tasks.get_task_dict([task])
-
-    for desc in fewshot_descriptions:
-        custom_task_dict = {k: CustomDescTask(v, desc) for k, v in task_dict.items()}
-
-        results = evaluator.evaluate(lm, custom_task_dict, True, num_fewshot, limit)
-
-        dumped = json.dumps(results, indent=2)
-
-        print('Description:', desc)
-        print(dumped)
-
-        # MAKE TABLE
-        from pytablewriter import MarkdownTableWriter
-
-        writer = MarkdownTableWriter()
-        writer.headers = ["Task", "Metric", "Value"]
-
-        values = []
-
-        for k, dic in results.items():
-            for m, v in dic.items():
-                values.append([k, m, '%.4f' % v])
-                k = ""
-        writer.value_matrix = values
-
-        print(writer.dumps())
-
-
-if __name__ == "__main__":
-    main()

scripts/get_prompts.py

@@ -9,7 +9,6 @@ for tname, Task in tasks.TASK_REGISTRY.items():#[('record', tasks.superglue.ReCo
     print('#', tname)
     docs = islice(task.validation_docs() if task.has_validation_docs() else task.test_docs(), ct)
     print()
-    print('**Zero-Shot Prompt**:', "\n```\n" + task.fewshot_description() + "\n```\n")
     for i in range(ct):
         print()
         doc = next(docs)

scripts/write_out.py

 import argparse
 import numpy as np
+import json
 import os
 import random
 from lm_eval import tasks
@@ -17,6 +18,7 @@ def parse_args():
     parser.add_argument('--num_fewshot', type=int, default=1)
     parser.add_argument('--seed', type=int, default=42)
     parser.add_argument('--num_examples', type=int, default=1)
+    parser.add_argument('--description_dict_path', default=None)
     return parser.parse_args()
 
 
@@ -29,6 +31,12 @@ def main():
     else:
         task_names = args.tasks.split(",")
     task_dict = tasks.get_task_dict(task_names)
+
+    description_dict = {}
+    if args.description_dict_path:
+        with open(args.description_dict_path, 'r') as f:
+            description_dict = json.load(f)
+
     os.makedirs(args.output_base_path, exist_ok=True)
     for task_name, task in task_dict.items():
         rnd = random.Random()
@@ -47,14 +55,16 @@ def main():
 
         docs = join_iters(iters)
 
+        description = description_dict[task_name] if description_dict and task_name in description_dict else ""
+
         with open(os.path.join(args.output_base_path, task_name), "w") as f:
             for i, doc in zip(range(args.num_examples), docs) if args.num_examples > 0 else enumerate(docs):
                 f.write(EXAMPLE_DIVIDER.format(i=i))
                 ctx = task.fewshot_context(
                     doc=doc,
-                    provide_description=args.provide_description,
                     num_fewshot=args.num_fewshot,
-                    rnd=rnd
+                    rnd=rnd,
+                    description=description
                 )
                 f.write(ctx + "\n")
 

tests/test_description_dict.py

+import random
+import lm_eval.tasks
+import lm_eval.models
+
+
+def test_description_dict():
+    seed = 42
+    num_examples = 1
+    task_names = ["hellaswag", "winogrande"]
+    description_dict = {
+        "hellaswag": "Label for the relevant action:\nSentences describing context, with an incomplete sentence trailing answer that plausibly completes the situation.",
+        "winogrande": "Winograd schema sentence including a either a ___ blank with a missing word, making the pronoun ambiguous, or the same with the word filled in.",
+    }
+
+    task_dict = lm_eval.tasks.get_task_dict(task_names)
+    for task_name, task in task_dict.items():
+        rnd = random.Random()
+        rnd.seed(seed)
+
+        if task.has_training_docs():
+            docs = task.training_docs()
+        elif set == "val" and task.has_validation_docs():
+            docs = task.validation_docs()
+        elif set == "test" and task.has_test_docs():
+            docs = task.test_docs()
+
+        description = (
+            description_dict[task_name]
+            if description_dict and task_name in description_dict
+            else ""
+        )
+
+        for _, doc in (
+            zip(range(num_examples), docs) if num_examples > 0 else enumerate(docs)
+        ):
+            ctx = task.fewshot_context(
+                doc=doc,
+                num_fewshot=1,
+                rnd=rnd,
+                description=description,
+            )
+            assert description in ctx

tests/test_evaluator.py

@@ -48,8 +48,22 @@ def test_evaluator(taskname, task_class):
     lm.loglikelihood_rolling = ll_perp_fn
 
     limit = 10
-    e1 = evaluator.evaluate(lm, task_dict, False, 0, limit, bootstrap_iters=10)
-    e2 = evaluator.evaluate(lm, task_dict, False, 0, limit, bootstrap_iters=10)
+    e1 = evaluator.evaluate(
+            lm=lm,
+            task_dict=task_dict,
+            num_fewshot=0,
+            limit=limit,
+            bootstrap_iters=10,
+            description_dict=None
+    )
+    e2 = evaluator.evaluate(
+            lm=lm,
+            task_dict=task_dict,
+            num_fewshot=0,
+            limit=limit,
+            bootstrap_iters=10,
+            description_dict=None
+    )
 
     # check that caching is working
     assert e1 == e2

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/test_version_stable.py

@@ -99,5 +99,13 @@ def test_versions_stable(taskname, task_class):
     lm.greedy_until = greedy_until
 
     limit = None
-    result = evaluator.evaluate(lm, task_dict, False, 0, limit, bootstrap_iters=10)
+    result = evaluator.evaluate(
+            lm=lm,
+            task_dict=task_dict,
+            num_fewshot=0,
+            limit=limit,
+            bootstrap_iters=10,
+            description_dict=None
+    )
+
     assert_target(f"{taskname}-v{task_class.VERSION}-res", result)