refactor(belle_sft): 重构处理数据集的逻辑

belle_sft.py

@@ -1,10 +1,39 @@
 from datasets import load_dataset
 from utils import process_dataset
 
+
+def build_prompt(query, history) -> str:
+    ret_prompt = ""
+    if len(history) > 0:
+        for his_query, his_response in history:
+            ret_prompt += f"<|user|> {his_query} <|system|> <bos>{his_response}<eos>\n"
+    if query is not None:
+        ret_prompt += f"<|user|> {query} <|system|> <bos>"
+    return ret_prompt
+
+
+def process_func(input_dict: dict):
+    conversations = input_dict["conversations"]
+    n = len(conversations) // 2
+    examples = []
+    
+    for i in range(n):
+        user_msg = conversations[2*i]["value"]
+        assistant_msg = conversations[2*i+1]["value"]
+        examples.append((user_msg, assistant_msg))
+    
+    content = {
+        "text": build_prompt(None, examples)
+    }
+    print(content)
+    return content
+
+
 if __name__ == "__main__":
     dataset = load_dataset("BelleGroup/train_3.5M_CN")
-    # process_dataset(
+    process_dataset(
-    #     dataset_dict=dataset,
+        dataset_dict=dataset,
-    #     output_subdir="belle_sft",
+        output_subdir="belle_sft",
-    #     max_chunk_size=5,
+        chunk_size=5,
-    # )
+        process_func=process_func,
+    )

utils.py

@@ -1,4 +1,4 @@
-from typing import List, Callable
+from typing import List, Callable, Union
 from datasets import DatasetDict
 from tokenizer import BpeTokenizer
 from tqdm import tqdm
@@ -64,7 +64,7 @@ def process_dataset(
     chunk_size: int = 1000000,
     split_name: str = "train",
     column_name: str = "text",
-    process_func: Callable[[dict], dict] = None,
+    process_func: Callable[[Union[dict, List[dict]]], dict] = None,
     normalization_func=comprehensive_normalization,
 ):
     train_dataset = dataset_dict[split_name]