Change to 0-based labels

Correspondingly out_of => scale_points

cross_domain_reviews.py

@@ -35,27 +35,17 @@ from datasets import concatenate_datasets, load_dataset
 class SubDataset:
     source: Any
     nick: str
-    out_of: int
+    scale_points: int
     get_review: Callable[[Any], Any]
     get_rating: Callable[[Any], Any]
 
 
-gi = operator.itemgetter
-
-
 def warn(msg):
     print(file=sys.stderr)
     print(f" ** Warning: {msg} **", file=sys.stderr)
     print(file=sys.stderr)
 
 
-def round_near(x, eps=0.001):
-    x_rnd = int(x + 0.5)
-    if abs(x_rnd - x) > eps:
-        warn("got {x_rnd} when rounding {x}")
-    return x_rnd
-
-
 @dataclass
 class SplitHFSrc:
     name: str
@@ -120,17 +110,42 @@ def int_or_drop(col):
     return inner
 
 
+gi = operator.itemgetter
+
+
+def round_near(x, eps=0.001):
+    x_rnd = int(x + 0.5)
+    if abs(x_rnd - x) > eps:
+        warn("got {x_rnd} when rounding {x}")
+    return x_rnd
+
+
+def dec(inner):
+    def wrap(x):
+        y = inner(x)
+        if y is None:
+            return None
+        res = y - 1
+        if res < 0:
+            warn(
+                "tried to convert 1-based index to 0-based index "
+                "but ended up with negative"
+            )
+        return res
+    return wrap
+
+
 DATASETS = [
-    SubDataset(SplitHFSrc("juliensimon/amazon-shoe-reviews"), "amazon-shoes", 5, gi("text"), lambda row: row["labels"] + 1),
+    SubDataset(SplitHFSrc("juliensimon/amazon-shoe-reviews"), "amazon-shoes", 5, gi("text"), gi("labels")),
     # TODO: Appears to be corrupt
     #SubDataset("florentgbelidji/edmunds-car-ratings", "car-ratings", 40, lambda row: row["Review"].strip(), lambda row: round_near(row["Rating"] * 8) - 7),
-    SubDataset(TrainOnlyHFSrc("florentgbelidji/car-reviews"), "car-ratings", 5, gi("Review"), gi("Rating")),
-    SubDataset(SplitHFSrc("codyburker/yelp_review_sampled"), "yelp", 5, gi("text"), gi("stars")),
-    SubDataset(SplitHFSrc("kkotkar1/course-reviews"), "course-reviews", 5, gi("review"), gi("label")),
-    SubDataset(TrainOnlyHFSrc("app_reviews"), "app-reviews", 5, gi("review"), gi("star")),
-    SubDataset(TrainOnlyHFSrc("LoganKells/amazon_product_reviews_video_games"), "amazon-games", 5, gi("reviewText"), lambda row: round_near(row["overall"]) + 1),
-    SubDataset(KaggleSrc("zynicide/wine-reviews", "winemag-data-130k-v2.csv"), "wine-reviews", 100, gi("description"), gi("points")),
-    SubDataset(KaggleSrc("sadmadlad/imdb-user-reviews", "Pulp Fiction/movieReviews.csv"), "imdb-user-reviews", 10, gi("Review"), int_or_drop("User's Rating out of 10")),
+    SubDataset(TrainOnlyHFSrc("florentgbelidji/car-reviews"), "car-ratings", 5, gi("Review"), dec(gi("Rating"))),
+    SubDataset(SplitHFSrc("codyburker/yelp_review_sampled"), "yelp", 5, gi("text"), dec(gi("stars"))),
+    SubDataset(SplitHFSrc("kkotkar1/course-reviews"), "course-reviews", 5, gi("review"), dec(gi("label"))),
+    SubDataset(TrainOnlyHFSrc("app_reviews"), "app-reviews", 5, gi("review"), dec(gi("star"))),
+    SubDataset(TrainOnlyHFSrc("LoganKells/amazon_product_reviews_video_games"), "amazon-games", 5, gi("reviewText"), lambda row: round_near(row["overall"])),
+    SubDataset(KaggleSrc("zynicide/wine-reviews", "winemag-data-130k-v2.csv"), "wine-reviews", 100, gi("description"), dec(gi("points"))),
+    SubDataset(KaggleSrc("sadmadlad/imdb-user-reviews", "Pulp Fiction/movieReviews.csv"), "imdb-user-reviews", 10, gi("Review"), dec(int_or_drop("User's Rating out of 10"))),
     # TODO: Unicode decoding error
     #SubDataset(KaggleSrc("arushchillar/disneyland-reviews", "DisneylandReviews.csv"), "disneyland-reviews", 5, gi("Review_Text"), gi("Rating")),
 ]
@@ -152,7 +167,7 @@ class CrossDomainReviews(datasets.GeneratorBasedBuilder):
             {
                 "text": datasets.Value("string"),
                 "rating": datasets.Value("uint8"),
-                "out_of": datasets.Value("uint8"),
+                "scale_points": datasets.Value("uint8"),
                 "dataset": datasets.Value("string")
             }
         )
@@ -202,26 +217,26 @@ class CrossDomainReviews(datasets.GeneratorBasedBuilder):
                 rating = ds_info.get_rating(row)
                 if rating is None:
                     continue
-                assert 1 <= rating <= ds_info.out_of, f"Expected {rating} (inclusively) between 1 and {ds_info.out_of}"
+                assert 0 <= rating < ds_info.scale_points, f"Expected {rating} in half-open (Python-style) range [0, {ds_info.scale_points})"
                 lowest = min(lowest, rating)
                 highest = max(highest, rating)
                 yield key, {
                     "text": review,
                     "rating": rating,
-                    "out_of": ds_info.out_of,
+                    "scale_points": ds_info.scale_points,
                     "dataset": ds_info.nick
                 }
                 key += 1
                 got += 1
                 if got >= 1000:
                     break
-            if lowest != 1:
+            if lowest != 0:
                 warn(
                     f"Lowest rating in {ds_info.nick} was {lowest}, "
-                    "would suppose it would be 1"
+                    "would suppose it would be 0"
                 )
-            if highest != ds_info.out_of:
+            if highest != ds_info.scale_points - 1:
                 warn(
                     f"Highest rating in {ds_info.nick} was {highest}, "
-                    f"would suppose it would be {ds_info.out_of}"
+                    f"would suppose it would be {ds_info.scale_points - 1}"
                 )