Add updated dataset for training llm.

dataset.py

@@ -0,0 +1,17 @@
+import json
+from torch.utils.data import Dataset
+
+
+class AudioDataset(Dataset):
+    def __init__(self, path):
+        with open(path, encoding='utf-8') as f:
+            self.dataset = json.load(f)
+
+    def __len__(self):
+        return len(self.dataset)
+
+    def __getitem__(self, idx):
+        text, audio, audio_path = self.dataset[idx]
+        # Format: [STOP][TEXT]<text prompt>[START]<audio tokens>[STOP]
+        res = f"[TEXT]{text}[START]{''.join(list(map(lambda x: f'[{x}]', audio)))}[STOP]"
+        return res