Model Fine-Tuning Times with Azure OpenAI vs. OpenAI Platform

Hello everyone,

I recently embarked on a project that involved fine-tuning a model using both the OpenAI platform and Azure OpenAI services. To my surprise, I encountered a significant discrepancy in the processing times between the two platforms.

On the OpenAI platform, the fine-tuning job was impressively swift, completing in just 20 minutes after submission. However, when I attempted to replicate the process on Azure OpenAI with identical training data, the job has markedly stretched beyond 12 hours and is yet to finish.

This disparity raises several questions and I’m curious to delve into the experiences of others. Have you utilized Azure OpenAI for model fine-tuning? If so, how do the completion times compare to those on the OpenAI platform for similar tasks? Moreover, are there specific configurations or best practices on Azure OpenAI that could optimize processing times?

I’d appreciate any insights, experiences, or advice on navigating these differences.



Azure’s costs for hosting fine-tuned models was prohibitively high for us to even consider it. Curious if you have calculated the costs as well?


Thank you for highlighting the cost considerations of using Azure OpenAI. Your insights prompt a reevaluation of our exploration into Azure as a backup solution. Given the significant processing time discrepancy and now the cost concerns you’ve raised, it seems more prudent to continue focusing on the OpenAI platform for the time being.

Hi, we are testing a fine-tuning project atm. Out of curiosity, how many training examples were included in your dateset? and what was the token usage in total? per example?

why don’t they just allow for serverless hot-swapping of loras?

paying for full uptime of a fine tuned deployment is absurd