Fix bugs

ultravox/data/datasets.py

@@ -401,11 +401,11 @@ def __iter__(self):
             actual_length += 1
             if actual_length == len(self) + 1:
                 warnings.warn(
-                    f"The presumed length {self._length} has been exceeded for split {self._dataset.split}. Make sure to update."
+                    f"The presumed length {self._length} has been exceeded for {self._config.name}:{self._args.split}. Make sure to update."
                 )
         if actual_length != len(self):
             warnings.warn(
-                f"Mismatch between presumed length ({self._length}) and actual length ({actual_length}) for split {self._dataset.split}. Make sure to update."
+                f"Mismatch between presumed length ({self._length}) and actual length ({actual_length}) for {self._config.name}:{self._args.split}. Make sure to update."
             )
 
     @abc.abstractmethod

ultravox/training/configs/meta_config.yaml

@@ -5,7 +5,7 @@ train_sets:
   - name: gigaspeech
 val_sets:
   - name: gigaspeech
-  - weight: 0.1
+  - weight: 0.01
 stop_strategy: "LAST_EXHAUSTED"
 
 train_on_inputs: False

ultravox/training/configs/release_config.yaml

@@ -12,7 +12,7 @@ loss_config:
 train_sets:
   - name: librispeech-clean-continuation
   - name: librispeech-other-continuation
-  - name: peoplespeech-continuation
+  - name: peoplespeech-clean-continuation
     weight: 8
   - name: common-voice-en-continuation
     weight: 8
@@ -56,7 +56,7 @@ train_sets:
     weight: 0.02
 
 # Temporarily remove heysquad_human from val_sets as it causes the training to fail.
-val_sets: ["peoplespeech"]
+val_sets: ["peoplespeech-clean"]
 
 batch_size: 24
 max_steps: 14400 # x8x24 = 2,764,800