subset2evaluate 1.0.18__tar.gz → 1.0.20__tar.gz

{subset2evaluate-1.0.18 → subset2evaluate-1.0.20}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: subset2evaluate
-Version: 1.0.18
+Version: 1.0.20
 Summary: Find informative examples to efficiently (human-)evaluate NLG models.
 Author-email: Vilém Zouhar <vilem.zouhar@gmail.com>
 License: MIT

{subset2evaluate-1.0.18 → subset2evaluate-1.0.20}/pyproject.toml

@@ -11,7 +11,7 @@ packages = ["subset2evaluate"]
 
 [project]
 name = "subset2evaluate"
-version = "1.0.18"
+version = "1.0.20"
 description = "Find informative examples to efficiently (human-)evaluate NLG models."
 license = {text = "MIT"}
 readme = "README.md"

{subset2evaluate-1.0.18 → subset2evaluate-1.0.20}/subset2evaluate/utils.py

@@ -840,6 +840,23 @@ def load_data_wmt(  # noqa: C901
                 # make sure it's positive
                 line["cost"] = (line["cost"] - cost_min) / (1 - cost_min)
 
+        # collect all humeval scores and try to deduce what happened there
+        humscores = [
+            model_v["human"] for line in data for model_v in line["scores"].values()
+        ]
+        if all(x <= 0 for x in humscores):
+            for line in data:
+                for model_v in line["scores"].values():
+                    model_v["human"] = max(0, min(100, 100 + 4 * model_v["human"]))
+        elif all(x >= 0 and x <= 1 for x in humscores):
+            for line in data:
+                for model_v in line["scores"].values():
+                    model_v["human"] = max(0, min(100, model_v["human"] * 100))
+        else:
+            for line in data:
+                for model_v in line["scores"].values():
+                    model_v["human"] = max(0, min(100, model_v["human"]))
+
         # this is min-max normalization
         if normalize and not binarize:
             _data_minmax_normalize(data)

{subset2evaluate-1.0.18 → subset2evaluate-1.0.20}/subset2evaluate.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: subset2evaluate
-Version: 1.0.18
+Version: 1.0.20
 Summary: Find informative examples to efficiently (human-)evaluate NLG models.
 Author-email: Vilém Zouhar <vilem.zouhar@gmail.com>
 License: MIT