Azure OpenAI微调报错：指定基础模型不支持的原因排查求助

运行的Python代码 #

import openai
from openai import cli
import time
import shutil
import json


openai.api_key = "*********************"
openai.api_base = "https://*********************" 
openai.api_type = 'azure'
openai.api_version = '2023-05-15' 
deployment_name ='*********************' 


training_file_name = 'training.jsonl'
validation_file_name = 'validation.jsonl'

# Samples data are fake
sample_data = [
               {"prompt": "Questa parte del testo e’ invece in italiano, perche’ Giuseppe Coco vive a Milano, codice postale 09576.", "completion": "[type: LOCATION, start: 36, end: 44, score: 0.85, type: PERSON, start: 54, end: 72, score: 0.85, type: LOCATION, start: 75, end: 81, score: 0.85]"},
               {"prompt": "In this fake document, we describe the ambarabacicicoco, of Alfred Johnson, who lives in Paris (France), the zip code is 21076, and his phone number is +32 475348723.", "completion": "[type: AU_TFN, start: 157, end: 166, score: 1.0, type: PERSON, start: 60, end: 74, score: 0.85, type: LOCATION, start: 89, end: 94, score: 0.85, type: LOCATION, start: 97, end: 103, score: 0.85, type: PHONE_NUMBER, start: 153, end: 166, score: 0.75]"},
               {"prompt": "This document is a fac simile", "completion": "[]"},
               {"prompt": "Here there are no PIIs", "completion": "[]"},
               {"prompt": "Questa parte del testo e’ invece in italiano, perche’ Giuseppe Coco vive a Milano, codice postale 09576.", "completion": "[type: LOCATION, start: 36, end: 44, score: 0.85, type: PERSON, start: 54, end: 72, score: 0.85, type: LOCATION, start: 75, end: 81, score: 0.85]"},
               {"prompt": "In this fake document, we describe the ambarabacicicoco, of Alfred Johnson, who lives in Paris (France), the zip code is 21076, and his phone number is +32 475348723.", "completion": "[type: AU_TFN, start: 157, end: 166, score: 1.0, type: PERSON, start: 60, end: 74, score: 0.85, type: LOCATION, start: 89, end: 94, score: 0.85, type: LOCATION, start: 97, end: 103, score: 0.85, type: PHONE_NUMBER, start: 153, end: 166, score: 0.75]"},
               {"prompt": "This document is a fac simile", "completion": "[]"},
               {"prompt": "Here there are no PIIs", "completion": "[]"},
               {"prompt": "10 August 2023", "completion": "[type: DATE_TIME, start: 0, end: 14, score: 0.85]"},
               {"prompt": "Marijn De Belie, Manu Brehmen (Deloitte Belastingconsulenten)", "completion": "[type: PERSON, start: 0, end: 15, score: 0.85, type: PERSON, start: 17, end: 29, score: 0.85]"},
               {"prompt": "The content expressed herein is based on the facts and assumptions you have provided us. We have assumed that these facts and assumptions are correct, complete and accurate.", "completion": "[]"},
               {"prompt": "This letter is solely for your benefit and may not be relied upon by anyone other than you.", "completion": "[]"},
               {"prompt": "Dear Mr. Mahieu,", "completion": "[type: PERSON, start: 9, end: 15, score: 0.85]"},
               {"prompt": "Since 1 January 2018, a capital reduction carried out in accordance with company law rules is partly imputed on the taxable reserves of the SPV", "completion": "[type: DATE_TIME, start: 6, end: 20, score: 0.85]"},
              ]

# Generate the training dataset file.
print(f'Generating the training file: {training_file_name}')
with open(training_file_name, 'w') as training_file:
    for entry in sample_data:
        json.dump(entry, training_file)
        training_file.write('\n')

# Copy the validation dataset file from the training dataset file.
# Typically, your training data and validation data should be mutually exclusive.
# For the purposes of this example, you use the same data.
print(f'Copying the training file to the validation file')
shutil.copy(training_file_name, validation_file_name)

def check_status(training_id, validation_id):
    train_status = openai.File.retrieve(training_id)["status"]
    valid_status = openai.File.retrieve(validation_id)["status"]
    print(f'Status (training_file | validation_file): {train_status} | {valid_status}')
    return (train_status, valid_status)

# Upload the training and validation dataset files to Azure OpenAI.
training_id = cli.FineTune._get_or_upload(training_file_name, True)
validation_id = cli.FineTune._get_or_upload(validation_file_name, True)

# Check the upload status of the training and validation dataset files.
(train_status, valid_status) = check_status(training_id, validation_id)

# Poll and display the upload status once per second until both files succeed or fail to upload.
while train_status not in ["succeeded", "failed"] or valid_status not in ["succeeded", "failed"]:
    time.sleep(1)
    (train_status, valid_status) = check_status(training_id, validation_id)

# This example defines a fine-tune job that creates a customized model based on curie,
# with just a single pass through the training data. The job also provides
# classification-specific metrics by using our validation data, at the end of that epoch.
create_args = {
    "training_file": training_id,
    "validation_file": validation_id,
    "model": "curie",
    "n_epochs": 1,
    "compute_classification_metrics": True,
    "classification_n_classes": 3
}

# Create the fine-tune job and retrieve the job ID and status from the response.
resp = openai.FineTune.create(**create_args)
job_id = resp["id"]
status = resp["status"]

# You can use the job ID to monitor the status of the fine-tune job.
# The fine-tune job might take some time to start and complete.
print(f'Fine-tuning model with job ID: {job_id}.')

# Get the status of our fine-tune job.
status = openai.FineTune.retrieve(id=job_id)["status"]

# If the job isn't yet done, poll it every 2 seconds.
if status not in ["succeeded", "failed"]:
    print(f'Job not in terminal status: {status}. Waiting.')
    while status not in ["succeeded", "failed"]:
        time.sleep(2)
        status = openai.FineTune.retrieve(id=job_id)["status"]
        print(f'Status: {status}')
else:
    print(f'Fine-tune job {job_id} finished with status: {status}')

# Check if there are other fine-tune jobs in the subscription. 
# Your fine-tune job might be queued, so this is helpful information to have
# if your fine-tune job hasn't yet started.
print('Checking other fine-tune jobs in the subscription.')
result = openai.FineTune.list()
print(f'Found {len(result)} fine-tune jobs.')

# Retrieve the name of the customized model from the fine-tune job.
result = openai.FineTune.retrieve(id=job_id)
if result["status"] == 'succeeded':
    model = result["fine_tuned_model"]

# Create the deployment for the customized model by using the standard scale type
# without specifying a scale capacity.
print(f'Creating a new deployment with model: {model}')
result = openai.Deployment.create(model=model, scale_settings={"scale_type":"standard", "capacity": None})

# Retrieve the deployment job ID from the results.
deployment_id = result["id"]

报错信息 #

openai.error.InvalidRequestError: The specified base model does not support fine-tuning.

背景情况 #

• 参考官方微调指南开发，服务部署在East US区域，了解到微调仅支持Central US但部署时无法找到该区域
• 尝试部署在North Central US区域，仍出现相同报错

报错原因分析 #

1. 区域支持限制：Azure OpenAI的微调功能仅对特定区域开放，当前官方支持的区域包含Central US。如果你的服务部署在未开放微调的区域（如East US、North Central US），即便指定了可微调的模型，也会触发该错误。部分订阅可能因权限或配额问题，无法看到Central US区域选项。
2. 模型名称格式错误：在Azure OpenAI环境中，微调时需使用Azure专属的模型标识符，而非OpenAI原生的curie。比如对应可微调的GPT-3.5系列模型，应使用gpt-35-turbo-instruct这类格式的名称，需确认订阅内可用的可微调模型列表。
3. 订阅权限/配额不足：若你的Azure订阅未获得微调功能的访问权限，或目标可微调模型的配额耗尽，也会出现该错误。需在Azure门户中检查OpenAI服务的权限配置和模型配额情况。

解决建议 #

• 确认区域权限：若无法找到Central US区域，联系Azure支持确认订阅是否有权限访问该区域，或等待微调功能在当前部署区域开放。
• 修正模型名称：在create_args的model参数中，替换为Azure OpenAI支持微调的模型标识符，比如gpt-35-turbo-instruct（需确认订阅内该模型可用）。
• 检查权限与配额：登录Azure门户，查看OpenAI服务的权限设置，确认已开启微调功能访问权限，同时检查目标模型的配额是否充足。

内容的提问来源于stack exchange，提问作者Gregory