changed chat dataset.

data.py

@@ -20,10 +20,11 @@ def dataWorker(q, dataset, encoding, batch_size, block_size,chat):
         for text in dataset:
             tokens = None
             if(chat):
-                txt = f"<|user|>{text['instruction']}"
-                if(text["input"] != None):
-                    txt += f"\n{text['input']}"
-                txt = txt + f"<|end|>\n<|assistant|>{text['output']}<|end|>"
+                txt=""
+                for msg in text['messages']:
+                    role = msg['role']
+                    content = msg['content']
+                    txt = txt + f"<|{role}|>{content}<|end|>"
                 tokens = [encoding.bos_token_id]+encoding.encode(txt)
             else:
                 tokens = [encoding.bos_token_id]+encoding.encode(text["text"])

train.py

@@ -23,8 +23,8 @@ hypr = {
     "encoding":    "gpt2",
     "dataset":     "HuggingFaceTB/smollm-corpus",
     "subset":      "cosmopedia-v2",
-    "chat_dataset": "yahma/alpaca-cleaned",
-    "chat_subset":  None,
+    "chat_dataset": "HuggingFaceTB/smoltalk",
+    "chat_subset":  "all",
     "half":         False,
 }