Merge pull request #88 from OpenBioLink/evaluation_cohere_model_outputs

improved evaluation function, simple checks

Author: KonstantinHebenstreit

libs/cot/cot/evaluate.py

@@ -10,6 +10,7 @@
 
 
 def search_regex(s: str, patterns: list, warn: bool) -> str:
+    """Searches a string for a list of regex patterns and returns the first found match."""
     # strip the string from whitespaces
     s = s.strip()
     for pattern in patterns:
@@ -80,20 +81,45 @@ def is_correct(type_: str, pred: str, gold: str, choices=None, warn=False) -> bo
         choices_dict = {"Yes": "True", "No": "False"}
         choices_keys = list(choices_dict.keys())
         choices_values = list(choices_dict.values())
+        choices_values_raw = (
+            choices_values  # in bool case, we need the raw values for the quick check
+        )
         keys_lower = [i.lower() for i in choices_dict.keys()]
         values_lower = [j.lower() for j in choices_dict.values()]
 
     # quick check if pred is in choices_dict
     if (
-        pred in choices_values
+        # We need to take the raw values here, as this is not regex
+        pred in choices_values_raw
         or pred in choices_keys
         or pred in keys_lower
         or pred in values_lower
     ):
+        # raise ValueError("not in choices_dict")
         is_correct = compare_pred_with_gold(pred, gold, choices_dict)
 
         return is_correct
 
+    # check if only one of the choices are part of the pred and report this as answer
+    # therefor search choice_value in pred and return if only one hit
+    hits = []
+    for value in choices_values:
+        # only check if length of value is smaller or same than pred
+        if len(value) <= len(pred):
+            # make value a group for regex
+            match = search_regex(
+                # "(" +  escape_special_characters(value) + ")", [escape_special_characters(pred)], warn
+                escape_special_characters(pred),
+                ["(" + value + ")"],
+                warn,
+            )
+            if match:
+                hits.append(match)
+            if len(hits) == 1:
+                pred = hits[0]
+                is_correct = compare_pred_with_gold(pred, gold, choices_dict)
+                return is_correct
+
     # if pred is not in choices_dict, we need to use regex
 
     # uppercase and lowercase is not important, as we will match the pattern case insensitive.

libs/cot/tests/integration_tests/test_evaluation_given_data.py

@@ -23,7 +23,7 @@ def test_evaluation_included_datasets():
 
     # compare with own calculation of the evaluation
     evaluation = collection.evaluate(overwrite=True, warn=False)
-    assert compare_nested_dict_float_values(evaluation, correct, 0.025)
+    assert compare_nested_dict_float_values(evaluation, correct, 0.021)
 
     # med_qa test set
     collection = Collection(["med_qa"], verbose=False)
@@ -49,9 +49,7 @@ def test_evaluation_included_datasets():
 
     # compare with own calculation of the evaluation
     evaluation = collection.evaluate(overwrite=True, warn=False)
-    assert compare_nested_dict_float_values(evaluation, correct, 0.001)
-    # was 0.00001 before. Got worse with individual answer sequences
-
+    assert compare_nested_dict_float_values(evaluation, correct, 1e-6)
 
     # medmc_qa validation set
     collection = Collection(["medmc_qa"], verbose=False)

libs/cot/tests/unit_tests/test_evaluate.py

@@ -145,38 +145,39 @@ def test_is_correct_multiple_answers():
 
 def test_predefined_correct_value():
     # med_qa
-    collection = Collection(["med_qa"], verbose=False)
-    collection = collection.select(
-        split="test", number_samples=10, random_samples=False
-    )
+    # collection = Collection(["med_qa"], verbose=False)
+    # collection = collection.select(
+    #     split="test", number_samples=10, random_samples=False
+    # )
 
-    collection2 = Collection(["med_qa"], verbose=False)
-    collection2 = collection2.select(
-        split="test", number_samples=10, random_samples=False
-    )
+    # collection2 = Collection(["med_qa"], verbose=False)
+    # collection2 = collection2.select(
+    #     split="test", number_samples=10, random_samples=False
+    # )
 
-    # only do evaluation on one of them, nothing should change
-    collection.evaluate(warn=False)
+    # # only do evaluation on one of them, nothing should change
+    # collection.evaluate(warn=False)
 
-    collection_json = collection.to_json()
-    collection2_json = collection2.to_json()
+    # collection_json = collection.to_json()
+    # collection2_json = collection2.to_json()
 
-    assert collection_json == collection2_json
+    # assert collection_json == collection2_json
 
     # pubmed_qa
     collection = Collection(["pubmed_qa"], verbose=False)
     collection = collection.select(
         split="train", number_samples=10, random_samples=False
     )
-    collection2 = Collection(["pubmed_qa"], verbose=False)
-    collection2 = collection2.select(
-        split="train", number_samples=10, random_samples=False
-    )
+    # collection2 = Collection(["pubmed_qa"], verbose=False)
+    # collection2 = collection2.select(
+    #     split="train", number_samples=10, random_samples=False
+    # )
+
+    collection_json = collection.to_json()
 
     # only do evaluation on one of them, nothing should change
-    collection.evaluate()
+    collection.evaluate(overwrite=False)
 
-    collection_json = collection.to_json()
-    collection2_json = collection2.to_json()
+    collection2_json = collection.to_json()
 
     assert collection_json == collection2_json