Revert training label normalization — normalize eval only

train_hf_jobs.py

@@ -276,7 +276,7 @@ def main():
                 batch["input_features"] = processor(
                     audio_array, sampling_rate=sr, return_tensors="pt"
                 ).input_features[0]
-                batch["labels"] = processor.tokenizer(normalizer(batch["phrase"])).input_ids
+                batch["labels"] = processor.tokenizer(batch["phrase"]).input_ids
                 return batch
 
             train_processed = train_dataset.map(preprocess, remove_columns=train_dataset.column_names)