This browser is no longer supported.
Upgrade to Microsoft Edge to take advantage of the latest features, security updates, and technical support.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(272, 53%, 36%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EAB%3C/text%3E%3C/svg%3E)
Hello Microsoft community,
We are excited with recently launched model called Phi-2. The SLM performance compared to LLMs ( <10 millions parameters) is impressive. We are excited to test the model but however we couldn't find any information on finetuning it. Is it possible to fine-tune the model for specific task like categorisation? If so, How can one fine tune phi-2 model and use it for inference?
@Abhishek Bhadoriya I think you are talking about announcements on this page. The current capabilities and limitations of the model are documented on ml.azure.com on the model catalog(preview) page. You should be able to view the same with your Azure ML workspace and deploy the same if the preview is turned on for your subscription.
Yes @romungi-MSFT , Thanks for responding back.
We have enabled the preview and able to deploy the model from the model catalog however I don't see any option to fine tune the model like phi-1.5 (Phi 1.5 has finetune feature available, just next to Evaluate). Therefore, I was hoping if there's a way to fine tune the model on azure and test it for specific tasks.
Yes, this is currently unavailable as the model hasn't been fine-tuned through reinforcement learning yet. If you intend to use the phi model for with finetuning you can use version 1.5
Thanks for the info
I do have one last question though, Do you have a rough estimate until when the finetuning would be available for phi-2 model? Any best guess?
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(48, 24%, 14%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EAD%3C/text%3E%3C/svg%3E)
Will Phi-2 only be available via Azure or are there plans to open source the model weights?
Thank you for publishing the model: https://huggingface.co/microsoft/phi-2
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(278.4, 77%, 36%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EMK%3C/text%3E%3C/svg%3E)
@romungi-MSFT I can see there is now a fine-tuning option available. I have tried to fine-tune Phi-2 on my data. However, for 160 samples, I ran it for 5 days and it was still running. I stopped the job and tried again, 2 days passed, and it is still running. Is that usually the case? How much do you think it should take? Thanks