最近领导要求我带一带他手下的研究生，课题是关于DAIL-SQL这个开源模型调试的。研究生说他已经完成了项目的搭建和gpt模型的调试。接下来要改成百度千帆的模型，却遇到了难题。
其实很简单，根据项目的框架走就可以了。正常部署环境过后（再次会遇到依赖安装不上，版本不匹配等诸多问题，按照报错排错修改即可）。根据readme生成数据模型之后。接下来就该调试ask_llm.py这一步开始了正式的模型修改了。
首先更改一下init方式
# 初始化百度千帆的api
init_qianfan(args.QIANFAN_ACCESS_KEY, args.QIANFAN_SECRET_KEY, args.model)然后查看到init方法是自己写的chatgpt的接口解析，我们就需要模仿人家的写法写一个百度千帆的，根据ERNIE_Speed_128K接口模型进行调试更改
https://cloud.baidu.com/doc/WENXINWORKSHOP/s/6ltgkzya5
之后得出了如下的脚本qianfan.py
import json.decoder
import qianfan
import os
from utils.enums import LLM
import time
def init_qianfan(QIANFAN_ACCESS_KEY, QIANFAN_SECRET_KEY,model):
os.environ["QIANFAN_ACCESS_KEY"] = QIANFAN_ACCESS_KEY
os.environ["QIANFAN_SECRET_KEY"] = QIANFAN_SECRET_KEY
# def init_qianfan(QIANFAN_ACCESS_KEY, QIANFAN_SECRET_KEY, model):
# qianfan.AccessKey(QIANFAN_ACCESS_KEY)
# qianfan.SecretKey(QIANFAN_SECRET_KEY)
# os.environ["QIANFAN_ACCESS_KEY"] = QIANFAN_ACCESS_KEY
# os.environ["QIANFAN_SECRET_KEY"] = QIANFAN_SECRET_KEY
# 处理单轮对话的completion任务
def ask_completion(model, batch, temperature):
completion = qianfan.Completion()
response = completion.do(
model=model,
prompt=batch, # 这是当前问题
temperature=temperature,
max_output_tokens=200, # 最大输出token数量，根据需要调整
top_p=1,
frequency_penalty=0,
presence_penalty=0,
stop=[";"]
)
# 提取response中的结果部分
response_clean = [response["result"]]
return dict(
response=response_clean,
prompt_tokens=response["usage"]["prompt_tokens"],
completion_tokens=response["usage"]["completion_tokens"],
total_tokens=response["usage"]["total_tokens"]
)
# 处理多轮对话的任务
def ask_chat(model, messages: list, temperature, n):
chat_completion = qianfan.ChatCompletion()
response = chat_completion.do(
model=model,
messages=messages, # messages 是带有历史对话的消息列表
temperature=temperature,
max_output_tokens=200 # 最大输出token数量，根据需要调整
)
# 提取返回的消息内容
response_clean = [response["result"]]
if n == 1:
response_clean = response_clean[0]
return dict(
response=response_clean,
prompt_tokens=response["usage"]["prompt_tokens"],
completion_tokens=response["usage"]["completion_tokens"],
total_tokens=response["usage"]["total_tokens"]
)
# 调用对话请i去函数，p判断ask_completion/ask_chat
def ask_llm(model: str, batch: list, temperature: float, n: int):
n_repeat = 0
while True:
try:
if model in LLM.TASK_COMPLETIONS: # completion任务
assert n == 1
response = ask_completion(model, batch, temperature)
elif model in LLM.TASK_CHAT: # chat任务
assert len(batch) == 1, "batch must be 1 in this mode"
messages = [{"role": "user", "content": batch[0]}]
response = ask_chat(model, messages, temperature, n)
response['response'] = [response['response']]
break
except json.decoder.JSONDecodeError:
n_repeat += 1
print(f"Repeat for the {n_repeat} times for JSONDecodeError", end="\n")
time.sleep(1)
continue
except Exception as e:
n_repeat += 1
print(f"Repeat for the {n_repeat} times for exception: {e}", end="\n")
time.sleep(1)
continue
return response接下来你如果运行会发现在ask_llm的时候异常会报错，这是因为异常不匹配，我没有细察qianfan的异常有那些，就直接抛出Exception
try:
res = ask_llm(args.model, batch, args.temperature, args.n)
except Exception as e:
print(f"The {i}-th question has too much tokens! Return \"SELECT\" instead")
# res = ""
res = {"response": [""], "total_tokens": 0}因为二者模型不一样，还需要对结果集进行数据的搜集更改，我这里没做处理，其实只需要一个简单的正则表达式就可以,在如下更改就可以了。
for sqls, db_id in zip(res["response"], cur_db_ids):
processed_sqls = []
for sql in sqls:
sql = " ".join(sql.replace("\n", " ").split())
sql = process_duplication(sql)
if sql.startswith("SELECT"):
pass
elif sql.startswith(" "):
sql = "SELECT" + sql
else:
sql = "SELECT " + sql
processed_sqls.append(sql)由于我没做更改，只得在后续生成的文件中进行数据提取，采用如下方法
def extract_sql(content):
# 匹配```sql开头 ```结尾
pattern = r'```sql(.*?)```'
sql_blocks = re.findall(pattern, content, re.DOTALL)
sql_statements = []
for sql in sql_blocks:
cleaned_sql = " ".join(sql.split())
sql_statements.append(cleaned_sql)
return sql_statements同时还需要再enmus脚本中添加对应的模型来识别
class LLM:
# openai LLMs
TEXT_DAVINCI_003 = "text-davinci-003"
CODE_DAVINCI_002 = "code-davinci-002"
GPT_35_TURBO = "gpt-3.5-turbo"
GPT_35_TURBO_0613 = "gpt-3.5-turbo-0613"
GPT_35_TURBO_16K = "gpt-3.5-turbo-16k"
GPT_35_TURBO_0301 = "gpt-3.5-turbo-0301"
GPT_4 = "gpt-4"
ERNIE_Speed_128K = 'ERNIE-Speed-128K'
# LLMs that use openai completion api
TASK_COMPLETIONS = [
TEXT_DAVINCI_003,
CODE_DAVINCI_002
]
# LLMs that use openai chat api
TASK_CHAT = [
GPT_35_TURBO,
GPT_35_TURBO_0613,
GPT_35_TURBO_16K,
GPT_35_TURBO_0301,
GPT_4,
ERNIE_Speed_128K
]脚本ask_llm输入的命令行修改为类似如下的方式
parser = argparse.ArgumentParser()
parser.add_argument("--question", type=str)
parser.add_argument("--QIANFAN_ACCESS_KEY", type=str)
parser.add_argument("--QIANFAN_SECRET_KEY", type=str)
# parser.add_argument("--openai_api_key", type=str)
# parser.add_argument("--openai_group_id", type=str, default="org-ktBefi7n9aK7sZjwc2R9G1Wo")
parser.add_argument("--model", type=str, choices=[LLM.TEXT_DAVINCI_003,
LLM.GPT_35_TURBO,
LLM.GPT_35_TURBO_0613,
# LLM.TONG_YI_QIAN_WEN,
LLM.GPT_35_TURBO_16K,
LLM.GPT_4,
LLM.ERNIE_Speed_128K],
default=LLM.ERNIE_Speed_128K)
parser.add_argument("--start_index", type=int, default=0)
parser.add_argument("--end_index", type=int, default=1000000)
parser.add_argument("--temperature", type=float, default=0) # qianfan (0, 1.0]
parser.add_argument("--mini_index_path", type=str, default="")
parser.add_argument("--batch_size", type=int, default=1)
parser.add_argument("--n", type=int, default=5, help="Size of self-consistent set")
parser.add_argument("--db_dir", type=str, default="dataset/spider/database")注意千帆的temperature是 (0, 1.0]和gpt的还不一样
至此已经完全修改完毕。