finetune添加system_prompt支持

cli_demo.py

@@ -9,10 +9,13 @@ import argparse
 import os
 import platform
 import shutil
+import json
+import json5
 from copy import deepcopy
 
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import AutoPeftModelForCausalLM
 from transformers.generation import GenerationConfig
 from transformers.trainer_utils import set_seed
 
@@ -40,8 +43,52 @@ Commands:
     :reset-conf         Reset generation config             重置生成配置
 '''
 
+TOOL_DESC = """{name_for_model}: 你可以调用该工具与 {name_for_human} API 进行交互。{name_for_human} API 有什么作用？{description_for_model} 参数列表：{parameters}"""
+
+REACT_INSTRUCTION = """请尽可能回答下列问题。您可以访问以下 API：
+
+{tools_text}
+
+使用以下格式回答问题：
+
+Question: 你需要回答的问题
+Thought: 你的思考过程
+Action: 要使用的操作，必须是 [{tools_name_text}] 其中之一
+Action Input: 操作的输入参数
+Observation: 操作的结果
+... (这些 Thought/Action/Action Input/Observation 可以是零次或重复多次)
+Thought: 你的最终思考过程
+Final Answer: 你的最终回答"""
+
+
+def build_react_instruction(functions: list[dict]):
+    tools_text = []
+    tools_name_text = []
+    for func_info in functions:
+        name = func_info.get("name", "")
+        name_m = func_info.get("name_for_model", name)
+        name_h = func_info.get("name_for_human", name)
+        desc = func_info.get("description", "")
+        desc_m = func_info.get("description_for_model", desc)
+        tool = TOOL_DESC.format(
+            name_for_model=name_m,
+            name_for_human=name_h,
+            description_for_model=desc_m,
+            parameters=json.dumps(func_info["parameters"], ensure_ascii=False),
+        )
+        tools_text.append(tool)
+        tools_name_text.append(name_m)
+    tools_text = "\n\n".join(tools_text)
+    tools_name_text = ", ".join(tools_name_text)
+    instruction = REACT_INSTRUCTION.format(
+        tools_text=tools_text,
+        tools_name_text=tools_name_text,
+    )
+    return instruction
+
 
 def _load_model_tokenizer(args):
+    model_path = args.model_path or args.checkpoint_path
     tokenizer = AutoTokenizer.from_pretrained(
         args.checkpoint_path, trust_remote_code=True, resume_download=True,
     )
@@ -49,9 +96,9 @@ def _load_model_tokenizer(args):
     if args.cpu_only:
         device_map = "cpu"
     else:
-        device_map = "auto"
+        device_map = "cuda"
 
-    model = AutoModelForCausalLM.from_pretrained(
+    model = AutoPeftModelForCausalLM.from_pretrained(
         args.checkpoint_path,
         device_map=device_map,
         trust_remote_code=True,
@@ -59,7 +106,7 @@ def _load_model_tokenizer(args):
     ).eval()
 
     config = GenerationConfig.from_pretrained(
-        args.checkpoint_path, trust_remote_code=True, resume_download=True,
+        model_path, trust_remote_code=True, resume_download=True,
     )
 
     return model, tokenizer, config
@@ -107,7 +154,13 @@ def main():
         description='QWen-Chat command-line interactive chat demo.')
     parser.add_argument("-c", "--checkpoint-path", type=str, default=DEFAULT_CKPT_PATH,
                         help="Checkpoint name or path, default to %(default)r")
+    parser.add_argument("-m", "--model-path", type=str, default=None,
+                        help="Model name or path, default to None")
     parser.add_argument("-s", "--seed", type=int, default=1234, help="Random seed")
+    parser.add_argument("-sf", "--system-prompt-file", type=str, default=None,
+                        help="System prompt file, default to None")
+    parser.add_argument("-fd", "--function-definition", type=str, default=None,
+                        help="Function definition file, should be json or json5, default to None")
     parser.add_argument("--cpu-only", action="store_true", help="Run demo with CPU only")
     args = parser.parse_args()
 
@@ -116,11 +169,24 @@ def main():
     model, tokenizer, config = _load_model_tokenizer(args)
     orig_gen_config = deepcopy(model.generation_config)
 
+    system_prompt = "You are a helpful assistant."
+    if args.system_prompt_file:
+        with open(args.system_prompt_file, 'r', encoding="utf-8") as f:
+            system_prompt = f.read()
+
+    function_prompt = None
+    if args.function_definition:
+        with open(args.function_definition, 'r', encoding="utf-8") as f:
+            functions = json5.load(f)
+            function_prompt = build_react_instruction(functions)
+
     _clear_screen()
     print(_WELCOME_MSG)
 
     seed = args.seed
 
+    is_first_msg = True
+
     while True:
         query = _get_input()
 
@@ -195,7 +261,11 @@ def main():
         # Run chat.
         set_seed(seed)
         try:
-            for response in model.chat_stream(tokenizer, query, history=history, generation_config=config):
+            prompt = query
+            if function_prompt:
+                prompt = f"{function_prompt}\n\nQuestion: {query}"
+
+            for response in model.chat_stream(tokenizer, prompt, history=history, system=system_prompt, generation_config=config):
                 _clear_screen()
                 print(f"\nUser: {query}")
                 print(f"\nQwen-Chat: {response}")

examples/function_call_finetune_examples.py

@@ -88,36 +88,37 @@ def main():
 
     example_functions = [
         {
-            "name_for_human": "Google Search",
-            "name_for_model": "google_search",
-            "description_for_model": "Google Search is a general search engine that can be used to access the internet,"
-            + " query encyclopedia knowledge, and stay informed about current events."
-            + " Format the arguments as a JSON object.",  # If you expect `Action Input` to be a JSON.
+            "name_for_human": "在线搜索",
+            "name_for_model": "search",
+            "description_for_model": "使用此工具可以搜索群内成员的信息，或者在互联网上搜索信息。Format the arguments as a JSON object.",  # If you expect `Action Input` to be a JSON.
             "parameters": [
                 {
-                    "name": "search_query",
-                    "description": "Search keywords or phrases",
-                    "required": True,  # Set to False if it is an optional parameter.
+                    "name": "keywords",
+                    "description": "需要搜索的关键词。",
+                    "required": True,
                     "schema": {"type": "string"},
                 },
-                # You can add more parameters to this `parameters` list if you wish.
             ],
         },
         {
-            "name_for_human": "Code Interpreter",
-            "name_for_model": "code_interpreter",
-            "description_for_model": "Code interpreter that can execute Python code."
-            + "Enclose the code within triple backticks (`)"
-            + " at the beginning and end of the code.",  # If you expect `Action Input` to be a Markdown code block.
+            "name_for_human": "天气信息",
+            "name_for_model": "get_weather",
+            "description_for_model": "查询某个地点的天气信息。Format the arguments as a JSON object.",
             "parameters": [
                 {
-                    "name": "code",
-                    "description": "Code to be executed",
+                    "name": "position",
+                    "description": "需要查询天气的地点。",
                     "required": True,
                     "schema": {"type": "string"},
                 },
             ],
         },
+        {
+            "name_for_human": "封禁用户",
+            "name_for_model": "ban_user",
+            "description_for_model": "在用户多次触犯道德规范时，使用此工具可以封禁用户。",
+            "parameters": [],
+        }
     ]
     example_instruction = build_react_instruction(example_functions)
 

finetune.py

@@ -221,7 +221,9 @@ class LazySupervisedDataset(Dataset):
         if i in self.cached_data_dict:
             return self.cached_data_dict[i]
         
-        ret = preprocess([self.raw_data[i]["conversations"]], self.tokenizer, self.max_len)
+        system_prompt: str = self.raw_data[i].get("system") or "You are a helpful assistant."
+
+        ret = preprocess([self.raw_data[i]["conversations"]], self.tokenizer, self.max_len, system_prompt)
         ret = dict(
             input_ids=ret["input_ids"][0],
             labels=ret["labels"][0],

finetune/finetune_lora_single_gpu.sh

@@ -4,7 +4,7 @@ export CUDA_DEVICE_MAX_CONNECTIONS=1
 MODEL="Qwen/Qwen-7B" # Set the path if you do not want to load from huggingface directly
 # ATTENTION: specify the path to your training data, which should be a json file consisting of a list of conversations.
 # See the section for finetuning in README for more information.
-DATA="path_to_data"
+DATA="../dataset/kurita.json"
 
 function usage() {
     echo '