Update README.md

README.md

@@ -14,6 +14,8 @@ the test split only.
 Each example contains one audio clip and one target attribute. The task is to
 answer whether that attribute increases or decreases over time.
 
+Code: [ynklab/vocalgrad](https://github.com/ynklab/vocalgrad).
+
 ## Task
 
 Given an audio clip and an attribute name, predict one of two labels:
@@ -116,12 +118,15 @@ Current category-to-attribute mapping:
 
 ## Evaluation
 
-Normalize model output to lowercase text. Accepted labels are:
+Normalize model output to lowercase text and match the following whole words
+using word boundaries:
 
-- `increase`
-- `decrease`
+- `increase` or `increases` -> `increase`
+- `decrease` or `decreases` -> `decrease`
 
-Any other output should be treated as invalid and counted as incorrect.
+Accept a prediction only when exactly one direction is matched. Outputs matching
+both directions or neither direction, missing responses, and failed generations
+are counted as incorrect.
 
 Main metric:
 
@@ -129,6 +134,9 @@ Main metric:
 accuracy = (# correct predictions) / (# total samples)
 ```
 
+The denominator includes all evaluated clips with valid ground-truth labels,
+including those with invalid or missing predictions.
+
 Recommended breakdowns:
 
 - overall
@@ -168,4 +176,4 @@ The audio clips in VocalGrad are transformed derivatives of VCTK source clips.
 VocalGrad applies controlled time-varying transformations to create benchmark
 examples for detecting whether a target perceptual attribute increases or
 decreases over time. The included metadata identifies the source clip ID and
-the transformation condition for each derived clip.
+the transformation condition for each derived clip.