merged changes

lm_eval/tasks/superglue.py

@@ -28,10 +28,10 @@ class BoolQ(HFTask):
         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']) 
+        return " " + yesno(doc['label']) 
 
     def construct_requests(self, doc, ctx):
 
@@ -156,12 +156,12 @@ class Copa(HFTask):
             "cause": "because",
             "effect": "therefore",
         }[doc["question"]]
-        return doc["premise"].strip()[:-1] + f" {connector} "
+        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)
+        return " " + self.convert_choice(correct_choice)
 
     def construct_requests(self, doc, ctx):
         choice1 = " " + self.convert_choice(doc["choice1"])
@@ -261,7 +261,7 @@ class ReCoRD(HFTask):
         return True
 
     def has_test_docs(self):
-        return True
+        return False
 
     def fewshot_description(self):
         # TODO: figure out actual description
@@ -322,6 +322,7 @@ class ReCoRD(HFTask):
         # - Evaluate the accuracy and token F1 PER EXAMPLE
         # - Average over all examples
         max_idx = np.argmax(np.array(results))
+
         prediction = doc["entities"][max_idx]
         gold_label_set = list(set(doc["answers"]))
         f1 = metric_max_over_ground_truths(squad_metrics.compute_f1, prediction, gold_label_set)

lm_eval/tasks/triviaqa.py

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

lm_eval/tasks/webqs.py

 from . common import HFTask
+from lm_eval.base import mean, rf
 
 class WebQs(HFTask):
     DATASET_PATH = "web_questions"
@@ -18,7 +19,6 @@ class WebQs(HFTask):
         return ""
 
     def doc_to_text(self, doc):
-        print(doc)
         return "Q: " + doc['question'] + '\nA:'
 
     def doc_to_target(self, doc):
@@ -27,47 +27,36 @@ class WebQs(HFTask):
         # TODO: make sure we're actually handling multi-answer correctly
         return " " + doc['answers'][0]
         
-    def construct_requests(self, doc, ctx):
-        """ Uses RequestFactory to construct Requests and returns an iterable of 
-        Requests which will be sent to the LM.
+    def _remove_prefixes(self, aliases):
+        # Optimization: Remove any alias that has a strict prefix elsewhere in the list
+        # we can do this because if the prefix is acceptable by isgreedy, we can stop looking
+        aliases.sort()
+        ret = [aliases[0]]
+        for alias in aliases[1:]:
+            if not alias.startswith(ret[-1]):
+                ret.append(alias)
 
-        :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`. 
-        """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        return ret
         
-    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.
-        """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+    def construct_requests(self, doc, ctx):
+        ret = []
+        for alias in self._remove_prefixes(doc['answers']):
+            _, is_prediction = rf.loglikelihood(ctx, " " + alias)
+            ret.append(is_prediction)
+        return ret
+
+    def process_results(self, doc, results):
+        return {
+            "acc": float(any(results))
+        }
 
     def aggregation(self):
-        """
-        :returns: {str: [float] -> float}
-            A dictionary where keys are the names of submetrics and values are 
-            functions that aggregate a list of metrics
-        """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        return {
+            "acc": mean,
+        }
 
     def higher_is_better(self):
-        """
-        :returns: {str: bool}
-            A dictionary where keys are the names of submetrics and values are 
-            whether a higher value of the submetric is better
+        return {
+            "acc": True
+        }
\ No newline at end of file
-        """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')

lm_eval/tasks/winogrande.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 HFTask, simple_accuracy_metric, yesno
+from . common import HFTask
+from lm_eval.base import rf, mean
+
+"""
+This evaluation of Winogrande uses partial evaluation as described by
+Trinh & Le in Simple Method for Commonsense Reasoning (2018).
+Reference: https://arxiv.org/abs/1806.02847
+"""
+
 
 class Winogrande(HFTask):
     DATASET_PATH = "winogrande"
@@ -17,35 +22,31 @@ class Winogrande(HFTask):
     def has_test_docs(self):
         return True
 
-    def training_docs(self):
-        if self.has_training_docs():
-            return self.data["train"]
-
-    def validation_docs(self):
-        if self.has_validation_docs():
-            return self.data["validation"]
-
-    def test_docs(self):
-        if self.has_test_docs():
-            return self.data["test"]
-
     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):
+        # Substitute the pronoun in the sentence with each candidate choice
+        # and ignore everything after.
+        pronoun_loc = doc["sentence"].index("_")
+        context1 = doc["sentence"][:pronoun_loc] + doc["option1"]
+        context2 = doc["sentence"][:pronoun_loc] + doc["option2"]
+        return context1, context2
+
+    @classmethod
+    def partial_target(cls, doc):
+        # The target is everything after the document specified pronoun.
+        pronoun_loc = doc["sentence"].index("_") + 1
+        return doc["sentence"][pronoun_loc:].strip()
+
     def doc_to_text(self, doc):
-        return doc['sentence']
+        context1, context2 = self.partial_context(doc)
+        return context1 + '\n' + context2 + '\n'
 
     def doc_to_target(self, doc):
-        text = doc['sentence']
-        answer_n = doc['answer']
-        if answer_n == '1':
-            answer = doc['option1']
-        elif answer_n == '2':
-            answer = doc['option2']
-        else:
-            raise ValueError("Winogrande from HF datasets contained an invalid answer key")
-        return text.replace("_", answer)
+        return self.partial_target(doc)
 
     def construct_requests(self, doc, ctx):
         """ Uses RequestFactory to construct Requests and returns an iterable of 
@@ -58,8 +59,11 @@ class Winogrande(HFTask):
             language description, as well as the few shot examples, and the question
             part of the document for `doc`. 
         """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        target = self.partial_target(doc)
+        context1, context2 = self.partial_context(doc)
+        ll_context1, _ = rf.loglikelihood(context1, " " + target)
+        ll_context2, _ = rf.loglikelihood(context2, " " + target)
+        return ll_context1, ll_context2
 
     def process_results(self, doc, results):
         """Take a single document and the LM results and evaluates, returning a 
@@ -71,8 +75,10 @@ class Winogrande(HFTask):
         :param results:
             The results of the requests created in construct_requests.
         """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        answer = int(doc["answer"]) - 1  # `- 1` b/c doc["answer"] ∈ {'1', '2'}
+        return {
+            "acc": np.argmax(results) == answer
+        }
 
     def aggregation(self):
         """
@@ -80,8 +86,9 @@ class Winogrande(HFTask):
             A dictionary where keys are the names of submetrics and values are 
             functions that aggregate a list of metrics
         """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        return {
+            "acc": mean
+        }
 
     def higher_is_better(self):
         """
@@ -89,5 +96,6 @@ class Winogrande(HFTask):
             A dictionary where keys are the names of submetrics and values are 
             whether a higher value of the submetric is better
         """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        return {
+            "acc": True
+        }

lm_eval/tasks/wsc273.py

-import json
+import numpy as np
 import random
-import os
-from lm_eval.base import Dataset
-from ..utils import sh
+from lm_eval.base import rf, mean
+from . common import HFTask
 
+"""
+NOTE: This evaluation of Winograd Schema Challenge is based on `partial evaluation`
+as described by Trinh & Le in Simple Method for Commonsense Reasoning (2018).
+See: https://arxiv.org/abs/1806.02847
+"""
+
+
+class WinogradSchemaChallenge273(HFTask):
+    DATASET_PATH = "winograd_wsc"
+    DATASET_NAME = "wsc273"
+
+    upper_pronouns = ["A", "An", "The", "She", "He",
+                      "It", "They", "My", "His", "Her", "Their"]
 
-class WinogradSchemaChallenge273(Dataset):    
     def __init__(self):
         super().__init__()
-
-    def download(self):
-        if not os.path.exists('data/wsc273'):
-            sh("""
-                mkdir -p data/wsc273 
-                wget https://git.cse.msu.edu/bakerb15/nlp-final-project/raw/master/Winogard/reproduce/commonsense_test/wsc273.json -O data/wsc273/wsc273.json
-                """)
+        self.data = self.__clean_data()
+
+    def __clean_data(self):
+        # The HF implementation of `wsc273` is not `partial evaluation` friendly.
+        data = []
+        for doc in self.data["test"]:
+            doc["text"] = doc["text"].replace("  ", " ")
+            doc["options"][0] = self.__normalize_option(doc["options"][0], doc)
+            doc["options"][1] = self.__normalize_option(doc["options"][1], doc)
+            data.append(doc)
+        return {"test": data}
+
+    def __normalize_option(self, option, doc):
+        # Append `'s` to possessive determiner based options.
+        if doc["pronoun"].lower() in ["my", "his", "her", "our", "their"]: 
+            option += "'s"
+        # Appropriately lowercase the pronoun in the option.
+        pronoun = option.split()[0]
+        start_of_sentence = doc["text"][doc['pronoun_loc'] - 2] == '.'
+        if not start_of_sentence and pronoun in self.upper_pronouns:
+            return option.replace(pronoun, pronoun.lower())
+        return option
 
     def has_training_docs(self):
         return False
@@ -25,60 +51,35 @@ class WinogradSchemaChallenge273(Dataset):
     def has_test_docs(self):
         return True
 
-    def training_docs(self):
-        return []
-
-    def validation_docs(self):
-        return []
-
-    def test_docs(self):
-        myjson = json.load(open('data/wsc273/wsc273.json'))
-        return self.load_doc(myjson)
+    def fewshot_examples(self, k):
+        # 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)
 
     def fewshot_description(self):
         # TODO: redo description
         return "Winograd schema sentence with correct continuation. True. Winograd schema sentence with incorrect continuation. False."
 
-    def load_doc(self, myjson):
-        docs = []
-        for i in range(0, 273 * 2, 2):
-            item1 = myjson[i]
-            item2 = myjson[i+1]
-
-            if item1['question_id'] != item2['question_id']:
-                raise ValueError("WSC273 has missing completion pair.")
-
-            question_id = item1['question_id']
+    @classmethod
+    def partial_context(cls, doc):
+        # Substitute the pronoun in the original text with each candidate 
+        # choice and ignore everything after. 
+        context1 = doc["text"][:doc["pronoun_loc"]] + doc["options"][0]
+        context2 = doc["text"][:doc["pronoun_loc"]] + doc["options"][1]
+        return context1, context2
 
-            if item1['correctness'] == True:
-                doc = {
-                    'id': question_id,
-                    'completions': {
-                        'T': item1['substitution'],
-                        'F': item2['substitution'],
-                    },
-                }
-                
-            if item2['correctness'] == True:
-                doc = {
-                    'id': question_id,
-                    'completions': {
-                        'F': item1['substitution'],
-                        'T': item2['substitution'],
-                    },
-                }
-
-            docs.append(doc)
- 
-        return docs
+    @classmethod
+    def partial_target(cls, doc):
+        # The target is everything after the document specified pronoun.
+        start_index = doc["pronoun_loc"] + len(doc["pronoun"])
+        return doc["text"][start_index:].strip()
 
     def doc_to_text(self, doc):
-        # TODO: implement
-        pass
+        context1, context2 = self.partial_context(doc)
+        return context1 + '\n' + context2 + '\n'
 
     def doc_to_target(self, doc):
-        # TODO: implement
-        pass
+        return self.partial_target(doc)
 
     def construct_requests(self, doc, ctx):
         """ Uses RequestFactory to construct Requests and returns an iterable of 
@@ -91,8 +92,11 @@ class WinogradSchemaChallenge273(Dataset):
             language description, as well as the few shot examples, and the question
             part of the document for `doc`. 
         """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        target = self.partial_target(doc)
+        context1, context2 = self.partial_context(doc)
+        ll_context1, _ = rf.loglikelihood(context1, " " + target)
+        ll_context2, _ = rf.loglikelihood(context2, " " + target)
+        return ll_context1, ll_context2
 
     def process_results(self, doc, results):
         """Take a single document and the LM results and evaluates, returning a 
@@ -104,8 +108,9 @@ class WinogradSchemaChallenge273(Dataset):
         :param results:
             The results of the requests created in construct_requests.
         """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        return {
+            "acc": np.argmax(results) == doc["label"]
+        }
 
     def aggregation(self):
         """
@@ -113,8 +118,9 @@ class WinogradSchemaChallenge273(Dataset):
             A dictionary where keys are the names of submetrics and values are 
             functions that aggregate a list of metrics
         """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        return {
+            "acc": mean
+        }
 
     def higher_is_better(self):
         """
@@ -122,5 +128,6 @@ class WinogradSchemaChallenge273(Dataset):
             A dictionary where keys are the names of submetrics and values are 
             whether a higher value of the submetric is better
         """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        return {
+            "acc": True
+        }

lm_eval/utils.py

@@ -29,3 +29,14 @@ def simple_parse_args_string(args_string):
 def join_iters(iters):
     for iter in iters:
         yield from iter
+
+
+def chunks(iter, n):
+    arr = []
+    for x in iter:
+        arr.append(x)
+        if len(arr) == n:
+            yield arr
+            arr = []
+    
+    if arr: yield arr
\ No newline at end of file

lm_eval/utils_stream.py

-import os
-from functools import reduce
-import operator
-from tqdm import tqdm
-import json
-
-
-class ExitCodeError(Exception):
-    pass
-
-
-def sh(x):
-    if os.system(x):
-        raise ExitCodeError()
-
-def ls(x):
-    return [x + '/' + fn for fn in os.listdir(x)]
-
-def lsr(x):
-    if os.path.isdir(x):
-        return reduce(operator.add, map(lsr, ls(x)), [])
-    else:
-        return [x]
-
-def fwrite(fname, content):
-    with open(fname, 'w') as fh:
-        fh.write(content)
-
-def fread(fname):
-    with open(fname) as fh:
-        return fh.read()
-
-class each:
-    def __init__(self, f):
-        self.f = f
-
-    def __rrshift__(self, other):
-        return list(map(self.f, other))
-
-class filt:
-    def __init__(self, f):
-        self.f = f
-
-    def __rrshift__(self, other):
-        return list(filter(self.f, other))
-
-class apply:
-    def __init__(self, f):
-        self.f = f
-
-    def __rrshift__(self, other):
-        return self.f(other)
-
-class one:
-    def __rrshift__(self, other):
-        try:
-            if isinstance(other, list): 
-                assert len(other) == 1
-                return other[0]
-            return next(other)
-        except:
-            return None
-
-class join:
-    def __init__(self, sep):
-        self.sep = sep
-
-    def __rrshift__(self, other):
-        if other is None:
-            return
-        try:
-            return self.sep.join(other)
-        except:
-            return None
-
-
-Y = object()
-
-def id(x):
-    return x
-
-class Reflective:
-    def __getattribute__(self, f):
-        def _fn(*args, **kwargs):
-            return lambda x: x.__getattribute__(f)(*args, **kwargs)
-        return _fn
-    
-    def __getitem__(self, a):
-        return lambda x: x[a]
-    
-    def __mul__(self, other):
-        if other == Y:
-            def _f(x, y=None):
-                if y == None:
-                    x, y = x
-                
-                return x * y
-            return  _f
-        
-        return lambda x: x * other
-    
-    def __rmul__(self, other):
-        if other == Y:
-            def _f(x, y=None):
-                if y == None:
-                    x, y = x
-                
-                return y * x
-            return  _f
-        
-        return lambda x: other * x
-    
-    def __add__(self, other):
-        if other == Y:
-            def _f(x, y=None):
-                if y == None:
-                    x, y = x
-                
-                return x + y
-            return  _f
-        
-        return lambda x: x + other
-    
-    def __radd__(self, other):
-        if other == Y:
-            def _f(x, y=None):
-                if y == None:
-                    x, y = x
-                
-                return y + x
-            return  _f
-        
-        return lambda x: other + x
-
-# (b -> a -> b) -> b -> [a] -> b
-def foldl(f, init, arr):
-    curr = init
-    for elem in arr:
-        curr = f(curr, elem)
-    return curr
-
-# (a -> b -> b) -> b -> [a] -> b
-def foldr(f, init, arr):
-    curr = init
-    for elem in arr[::-1]:
-        curr = f(elem, curr)
-    return curr
-
-
-def comp(*fs):
-    if len(fs) == 1:
-        return fs[0]
-    
-    def _f(x):
-        for f in fs[::-1]:
-            x = f(x)
-    
-        return x
-    return _f
-
-
-X = Reflective()

main.py

@@ -4,9 +4,11 @@ import numpy as np
 import random
 import itertools
 import collections
+import logging
 
-from lm_eval import models, tasks
+from lm_eval import models, tasks, evaluator, base
 
+logging.getLogger("openai").setLevel(logging.WARNING)
 
 def parse_args():
     parser = argparse.ArgumentParser()
@@ -18,102 +20,26 @@ def parse_args():
     parser.add_argument('--seed', type=int, default=1234)
     parser.add_argument('--output_path', default=None)
     parser.add_argument('--limit', type=int, default=None)
+    parser.add_argument('--cache', action="store_true")
     return parser.parse_args()
 
 def main():
+
     args = parse_args()
     random.seed(args.seed)
     np.random.seed(args.seed)
 
     lm = models.get_model(args.model).create_from_arg_string(args.model_args)
+
+    if args.cache:
+        lm = base.CachingLM(lm, 'lm_cache/' + args.model + '_' + args.model_args.replace('=', '-').replace(',', '_') + '.db')
     if args.tasks == "all_tasks":
         task_names = tasks.ALL_TASKS
     else:
         task_names = args.tasks.split(",")
     task_dict = tasks.get_task_dict(task_names)
 
-    task_dict_items = []
-    for name, task in task_dict.items():
-        if task.has_validation_docs():
-            task_dict_items.append((name, task, 'validation'))
-        elif task.has_test_docs():
-            task_dict_items.append((name, task, 'test'))
-        elif task.has_training_docs():
-            task_dict_items.append((name, task, 'training'))
-        
-    results = collections.defaultdict(dict)
-
-    requests = collections.defaultdict(list)
-    requests_origin = collections.defaultdict(list)
-
-    # if we ever run into issues where the eval tasks don't fit in memory and we can't afford a machine with bigger memory,
-    # we can always modify this plumbing to support that, but i didn't want to include it just yet because overengineering is bad
-    # (or we could make it write the requests to disk and then read them back out again - probably using an sqlite db because of all the moving parts we have
-
-    # TODO: we need unit tests & sanity checks or something to ensure that the return of `validation_docs` is stable
-
-    docs = {}
-
-    # get lists of each type of requeste
-    for task_name, task, dset in task_dict_items:
-        if dset == 'training':
-            temp = task.training_docs()
-        elif dset == 'test':
-            temp = task.test_docs()
-        else:
-            temp = task.validation_docs()
-
-        for doc_id, doc in enumerate(itertools.islice(temp, 0, args.limit)):
-            docs[(task_name, doc_id)] = doc
-
-            ctx = task.fewshot_context(
-                doc=doc,
-                provide_description=args.provide_description,
-                num_fewshot=args.num_fewshot,
-            )
-
-            reqs = task.construct_requests(doc, ctx)
-
-            for i, req in enumerate(reqs):
-                requests[req.type].append(req)
-                # i: index in requests for a single task instance
-                # doc_id: unique id that we can get back to a doc using `docs`
-                requests_origin[req.type].append((i, task_name, doc, doc_id))
-
-    # all responses for each (task, doc)
-    process_res_queue = collections.defaultdict(list)
-
-    # execute each type of request
-    for reqtype, reqs in requests.items():
-        # TODO: right now, this code runs multiple seperate LM requests for multiple Requests differing
-        # only in index. We could implement some kind of caching, but that would be more of a bandaid
-        # solution. we could also implement some kind of autogrouping here; they should end up next to each other.
-
-        resps = getattr(lm, reqtype)([req.args for req in reqs])
-
-        resps = [x if req.index is None else x[req.index] for x, req in zip(resps, reqs)]
-
-        for resp, (i, task_name, doc, doc_id) in zip(resps, requests_origin[reqtype]):
-            process_res_queue[(task_name, doc_id)].append((i, resp))
-    
-    vals = collections.defaultdict(list)
-
-    # unpack results and sort back in order and return control to Task
-    for (task_name, doc_id), requests in process_res_queue.items():
-        requests.sort(key=lambda x: x[0])
-        requests = [x[1] for x in requests]
-
-        task = task_dict[task_name]
-        doc = docs[(task_name, doc_id)]
-
-        metrics = task.process_results(doc, requests)
-        for metric, value in metrics.items():
-            vals[(task_name, metric)].append(value)
-    
-    # aggregate results
-    for (task_name, metric), items in vals.items():
-        task = task_dict[task_name]
-        results[task_name][metric] = task.aggregation()[metric](items)
+    results = evaluator.evaluate(lm, task_dict, args.provide_description, args.num_fewshot, args.limit)
 
     dumped = json.dumps(results, indent=2)
     print(dumped)

requirements.txt

@@ -5,3 +5,4 @@ click>=7.1
 scikit-learn>=0.24.1
 torch>=1.7
 transformers>=4.1
+sqlitedict==1.6.0
\ No newline at end of file

scripts/make_table.py

+from lm_eval import tasks
+from pytablewriter import MarkdownTableWriter
+
+writer = MarkdownTableWriter()
+writer.headers = ["Task Name", "Train", "Val", "Test", "Metrics"]
+
+values = []
+
+def chk(tf):
+    if tf:
+        return '✓'
+    else:
+        return ' '
+
+for tname, Task in tasks.TASK_REGISTRY.items():
+    task = Task()
+
+    values.append([tname,chk(task.has_training_docs()),chk(task.has_validation_docs()),chk(task.has_test_docs()),', '.join(task.aggregation().keys())])
+
+writer.value_matrix = values
+
+print(writer.dumps())
\ No newline at end of file

tests/test_evaluator.py

+import lm_eval.tasks as tasks
+import lm_eval.models as models
+import lm_eval.evaluator as evaluator
+import random
+import pytest
+
+
+# TODO: more fine grained unit tests rather than this big honking integration
+# test once we break evaluator into smaller, more manageable pieces
+
+@pytest.mark.parametrize("taskname,Task", tasks.TASK_REGISTRY.items())
+def test_evaluator(taskname, Task):
+    task_dict = tasks.get_task_dict([taskname])
+    lm = models.get_model('dummy')()
+
+    def ll_fn(reqs):
+        for ctx, cont in reqs:
+            # space convention
+            assert ctx[-1] != ' '
+            assert cont[0] == ' ' or ctx[-1] == '\n'
+        
+        res = []
+        
+        random.seed(42)
+        for _ in reqs:
+            res.append((-random.random(), False))
+
+        return res
+        
+
+    lm.loglikelihood = ll_fn
+    evaluator.evaluate(lm, task_dict, False, 0, 10)
\ No newline at end of file

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")
+    (ll_dog, ig_dog), (ll_cat, ig_cat) = gpt2.loglikelihood([
+        ('The quick brown fox jumps over the lazy', ' dog'),
+        ('The quick brown fox jumps over the lazy', ' cat'),
+    ])
+
+    assert ll_dog > ll_cat
+    assert not ig_cat
+

tests/test_all_sanitycheck.py → tests/test_tasks.py

 import lm_eval.tasks as tasks
 import lm_eval.base as base
-from unittest.mock import MagicMock
 from itertools import islice
 import pytest
 
@@ -8,10 +7,10 @@ import pytest
 @pytest.mark.parametrize("taskname,Task", tasks.TASK_REGISTRY.items())
 def test_basic_interface(taskname, Task):
     print('Evaluating task', taskname)
-    dl = Task.download
-    Task.download = MagicMock()
+    #dl = Task.download
+    #Task.download = MagicMock()
     task = Task()
-    Task.download = dl
+    #Task.download = dl
 
     assert task.has_training_docs() in [True, False]
     assert task.has_validation_docs() in [True, False]
@@ -44,8 +43,12 @@ def test_documents_and_requests(taskname, Task):
             assert isinstance(txt, str)
             assert isinstance(tgt, str)
             
+            # space convention
+            assert txt[-1] != ' '
+            assert tgt[0] == ' ' or txt[-1] == '\n'
+
             reqs = task.construct_requests(doc, txt)
 
-            # todo: mock lm by pluggin what's currently in main.py in here
+            # todo: mock lm after refactoring evaluator.py to not be a mess
             for req in reqs:
                 assert isinstance(req, base.Request)