Fine-tuning 实战

微调通过输入给大模型更多的样本数据来改进模型的表现,使其在一些特定的任务上输出更好的结果。

使用场景:

Fine-tuning 日志分析专家

使用真实的已标注数据进行 Fine-tuning,使模型具备判断日志故障优先级的能力

📎 此处原稿为飞书表格附件,Markdown 导出时未包含该表格,内容暂缺。 需要将表格处理成如下的格式:

执行:

import csv
import json
import os


def convert_csv_to_jsonl(csv_file_path, jsonl_file_path):
    with open(csv_file_path, mode="r", encoding="utf-8") as csv_file:
        csv_reader = csv.DictReader(csv_file)

        with open(jsonl_file_path, mode="w", encoding="utf-8") as jsonl_file:
            for row in csv_reader:
                log = row["log"]
                priority = row["优先级"]

                jsonl_entry = {
                    "messages": [
                        {
                            "role": "system",
                            "content": "你现在是一个日志告警专家,请根据日志内容去识别紧急程度,直接输出 P0 P1 或 P2,输出:",
                        },
                        {"role": "user", "content": log},
                        {"role": "assistant", "content": priority},
                    ]
                }

                jsonl_file.write(json.dumps(jsonl_entry, ensure_ascii=False) + "\n")


csv_file_path = os.path.join(os.path.dirname(__file__), "data", "log.csv")
print(f"csv_file_path:{csv_file_path}")
jsonl_file_path = "log.jsonl"
print(f"jsonl_file_path:{jsonl_file_path}")
convert_csv_to_jsonl(csv_file_path, jsonl_file_path)

lesson/module_4/demo_6/process_data.py

Fine-tuning 数据格式:jsonl

包含 system prompt、user、assistant每行代表一条日志标注数据

# 02finetune.py
from openai import OpenAI

client = OpenAI()

# 上传微调数据
file_name = client.files.create(file=open("log.jsonl", "rb"), purpose="fine-tune")
file_id = file_name.id
print("file_id is: ", file_id)

# 创建微调任务
finetune_job = client.fine_tuning.jobs.create(
    training_file=file_id, model="gpt-4o-mini-2024-07-18"
)
job_id = finetune_job.id
print("job_id is: ", job_id)

流程:

注意:

Fine-tuning 日志分析专家-推理

# 04chat.py
from openai import OpenAI

client = OpenAI()

model_id = "ft:gpt-4o-mini-2024-07-18:he3::9tbKC81t"

completion = client.chat.completions.create(
    model=model_id,
    messages=[
        {
            "role": "system",
            "content": "你现在是一个日志告警专家,请根据日志内容去识别紧急程度,直接输出 P1 P2 或 P3,输出:",
        },
        {"role": "user", "content": "Disk I/O error"},
    ],
)

print("ChatGPT resp: ", completion.choices[0].message.content)