OpenAI 已经支持在它的模型上进行微调,包括 gpt-3.5-turbo,gpt-4。在这里,把如何微调 OpenAI 的 GPT-3.5-turbo-1106的过程记录下来。
一、准备数据集
微调任何人工智能模型的第一步是准备训练数据集。对于我们的示例,我们将使用名为 的 CSV 文件。
此例子中,translate.csv
其中包含与一些游戏领域的中文翻译内容。 origin
为原文,target
为译文
然后,我们将此 CSV 文件转换为更适合训练 AI 模型的 JSONL(JSON Lines)
格式。下面的 Python 脚本读取 CSV 文件并将其转换为 JSONL 格式:
python复制代码
import json
import pandas as pd
DEFAULT_SYSTEM_PROMPT = '把在>>>和<<<中的内容翻译成[[和]]中的语言 '
def get_example(language, origin, target):
return {
"messages": [
{"role": "system", "content": DEFAULT_SYSTEM_PROMPT},
{"role": "user", "content": f'[[{language}]], >>>{origin}<<<'},
{"role": "assistant", "content": target},
]
}
if __name__ == "__main__":
df = pd.read_csv("translate.csv")
with open("train.jsonl", "w", encoding="utf8") as f:
for i, row in list(df.iterrows()):
origin = row["origin"]
target = row["target"]
print(origin)
example = get_example('en', origin, target)
example_str = json.dumps(example,ensure_ascii=False)
f.write(example_str + "n")
生成的jsonl中的内容类似
具体每一个行类似
二、运行微调
一旦我们准备好训练数据集,我们就可以使用 OpenAI 的 API 继续微调我们的模型:
1、安装openai 包
注意:本文的例子需要 openai 版本 > 1.1.0 的,如果太老旧了,请更新
复制代码pip install openai
2、执行微调命令
python复制代码
import json
from time import sleep
from openai import OpenAI
import os
os.environ['OPENAI_API_KEY']="sk-7Vl54m90xxxxxxxxxxxxxxxxxxxxxxx"
client = OpenAI()
#client = OpenAI(api_key="sk-7Vl54m90xxxxxxxxxxxxx") #或者通过参数传入类似
model_name = 'gpt-3.5-turbo-1106'
#监控任务完成状态
def wait_untill_done(job_id):
events = {}
while True:
response = client.fine_tuning.jobs.list_events(fine_tuning_job_id=job_id, limit=10)
print('fine tuning, waiting for ...')
# collect all events
for event in response.data:
if "data" in event and event.data:
print(event.data)
events[event.data["step"]] = event.data["train_loss"]
messages = [it.message for it in response.data]
for m in messages:
if m.startswith("New fine-tuned model created: "):
return m.split("created: ")[1], events
sleep(10)
if __name__ == "__main__":
response = client.files.create(file=open("train.jsonl", "rb"), purpose="fine-tune")
uploaded_id = response.id
print('uploaded_id=', uploaded_id )
print("Dataset is uploaded")
print("Sleep 10 seconds...")
sleep(10) # wait until dataset would be prepared
response = client.fine_tuning.jobs.create(training_file=uploaded_id,model=model_name)
ft_job_id = response.id
print("Fine-tune job is started, job_id = ",ft_job_id)
new_model_name, events = wait_untill_done(ft_job_id)
with open("new_model_name.txt", "w") as fp:
fp.write(new_model_name)
print("Fine-tune job is success, new model name = ",new_model_name)
执行成功后,可以看到类似这样的内容
erlang复制代码
uploaded_id= file-3KzFOCxKqfZTZe89m1I40wgA
Dataset is uploaded
Sleep 30 seconds...
Fine-tune job is started, job_id = ftjob-PiqWqQ6BDPbB9hCELN2B6MbL
fine tuning, waiting for ...
fine tuning, waiting for ...
fine tuning, waiting for ...
fine tuning, waiting for ...
fine tuning, waiting for ...
fine tuning, waiting for ...
......
Fine-tune job is success, new model name = ft:gpt-3.5-turbo-1106:personal::8LqhuNgA
在执行的过程中,也可以上 platform上platform.openai.com/finetune,可以看到类似这样的内容。
3、记住微调后的model名
成功之后,可以看到输出 Fine-tune job is success, new model name = ft:gpt-3.5-turbo-1106:personal::8LqhuNgA
,记住这个新的model name。
如果不小心关掉了,也可以在 platform上查看。如
三、使用微调好的模型
成功微调我们的模型后,我们现在可以使用它根据用户输入生成响应:
python复制代码
from openai import OpenAI
import os
os.environ['OPENAI_API_KEY']="sk-7Vl54m90xxxxxxxxxxxxxxxxxxxxxxx"
client = OpenAI()
#client = OpenAI(api_key="sk-7Vl54m90xxxxxxxxxxxxx") #或者通过参数传入类似
response = client.chat.completions.create(
model="ft:gpt-3.5-turbo-1106:personal::8LqhuNgA", ##此处为上面微调好的新model
messages=[
{"role": "system", "content": "你是一个语言专家,把在>>>和<<<中的内容翻译成[[和]]中的语言 "},
{"role": "user", "content": "[[en]],>>>选择目标友方英雄开始施法,一段时间后传送至目标位置 施法期间右方英雄获得护盾值,并和{Hero_149}获得伤害减免 传送完成后{Hero_149}增加移动速度,自身周围一定范围内右方英雄获得物理防御和魔法防御<<<"}
]
)
print(response.choices[0].message.content)
得到这样的结果
css复制代码
selects a teammate and starts channeling, then teleports to the target after a while. While channeling, he grants a shield to heroes to the right and damage reduction to himself. After teleporting, his Movement Speed is increased, and heroes in range to the right gain Physical Defense and Magical Defense.