Upload dan-chat-advanced.py

dan-chat-advanced.py

@@ -20,6 +20,7 @@ turn_separator = "\n"
 system_prefix = "<|im_start|>system\n"
 user_prefix = "<|im_start|>user\n"
 assistant_prefix = "<|im_start|>assistant\n"
+tool_prefix = "<|im_start|>tool\n"
 
 class DanChatMLPromptTokenizingStrategy(PromptTokenizingStrategy):
     def __init__(self, prompter, tokenizer, train_on_inputs, sequence_len, *args, **kwargs):
@@ -41,12 +42,19 @@ class DanChatMLPromptTokenizingStrategy(PromptTokenizingStrategy):
         not_first_turn = False
         
         for role, message, loss, prefix in prompt_parts:
-            prefix = prefix or ""
-            message = prefix + message
+            # If prefix is not defined, set it to an empty string
+            if prefix is None:
+                prefix = ""
             
-            if role in ["system", "user", "human"]:
-                role_prefix = system_prefix if role == "system" else user_prefix
-                res = self._tokenize_with_turn(role_prefix, message, not_first_turn)
+            if role in ["system", "user", "human", "tool"]:
+                # Set the role prefix based on the role
+                if role == "system":
+                    role_prefix = system_prefix
+                elif role == "user" or role == "human":
+                    role_prefix = user_prefix
+                elif role == "tool":
+                    role_prefix = tool_prefix
+                res = self._tokenize_with_turn(role_prefix, prefix + message, not_first_turn)
                 labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
             
             elif role in ["model", "gpt"]: