refactor: SFT 统一 messages 格式 + ChatML 纯 jinja 渲染

scripts/cache_h5.py

@@ -34,8 +34,8 @@ def main():
     parser.add_argument(
         "-t",
         "--tokenizer",
-        default="./tokenizer.json",
-        help="Tokenizer path (default: ./tokenizer.json)",
+        default="./tokenizer",
+        help="Tokenizer dir (default: ./tokenizer)",
     )
     parser.add_argument(
         "-s",

scripts/supervised_finetuning/sft_alpaca_gpt4.py

@@ -6,10 +6,13 @@ def process_func(input_dict: dict):
     instruction = input_dict["instruction"]
     inp = input_dict.get("input", "")
     if inp:
-        query = instruction + "\n" + inp
+        content = instruction + "\n" + inp
     else:
-        query = instruction
-    return {"query": query, "response": input_dict["output"]}
+        content = instruction
+    return {"messages": [
+        {"role": "user", "content": content},
+        {"role": "assistant", "content": input_dict["output"]},
+    ]}
 
 
 if __name__ == "__main__":

scripts/supervised_finetuning/sft_firefly-1.1m-rephrased.py

@@ -3,7 +3,10 @@ from pipeline import export_dataset
 
 
 def process_func(input_dict: dict):
-    return {"query": input_dict["instruction"], "response": input_dict["output"]}
+    return {"messages": [
+        {"role": "user", "content": input_dict["instruction"]},
+        {"role": "assistant", "content": input_dict["output"]},
+    ]}
 
 
 if __name__ == "__main__":

scripts/supervised_finetuning/sft_magicoder.py

@@ -3,7 +3,10 @@ from pipeline import export_dataset
 
 
 def process_func(input_dict: dict):
-    return {"query": input_dict["instruction"], "response": input_dict["response"]}
+    return {"messages": [
+        {"role": "user", "content": input_dict["instruction"]},
+        {"role": "assistant", "content": input_dict["response"]},
+    ]}
 
 
 if __name__ == "__main__":

scripts/supervised_finetuning/sft_metamathqa.py

@@ -3,7 +3,10 @@ from pipeline import export_dataset
 
 
 def process_func(sample: dict) -> dict:
-    return {"query": sample["query"], "response": sample["response"]}
+    return {"messages": [
+        {"role": "user", "content": sample["query"]},
+        {"role": "assistant", "content": sample["response"]},
+    ]}
 
 
 if __name__ == "__main__":

scripts/supervised_finetuning/sft_openhermes.py

@@ -2,13 +2,30 @@ from datasets import load_dataset
 from pipeline import export_dataset
 
 
+ROLE_MAP = {"system": "system", "human": "user", "gpt": "assistant"}
+
+
 def process_func(input_dict: dict):
     conversations = input_dict["conversations"]
+
+    system_msgs = []
+    idx = 0
+    if conversations and conversations[0]["from"] == "system":
+        system_msgs.append({
+            "role": "system",
+            "content": conversations[0]["value"],
+        })
+        idx = 1
+
     examples = []
-    for i in range(0, len(conversations) - 1, 2):
-        user_msg = conversations[i]["value"]
-        assistant_msg = conversations[i + 1]["value"]
-        examples.append({"query": user_msg, "response": assistant_msg})
+    for i in range(idx, len(conversations) - 1, 2):
+        user_msg = conversations[i]
+        assistant_msg = conversations[i + 1]
+        messages = system_msgs + [
+            {"role": ROLE_MAP[user_msg["from"]], "content": user_msg["value"]},
+            {"role": ROLE_MAP[assistant_msg["from"]], "content": assistant_msg["value"]},
+        ]
+        examples.append({"messages": messages})
     return examples