Add Instructions for Running Polytropon in README

[kadarsh22]

Hello,

Could the README file be updated to include clear instructions on how to run Polytropon? It would be helpful for new users to have a section that outlines the steps required to get Polytropon up and running.

[pclucas14]

Hi,

Thanks for reaching out. Can you give me a bit of context regarding what you are trying to do ?

[kadarsh22]

I was trying to run it and recreate the results in the paper, but not much information is given on how to run it.

[shuishen112]

I see. Thanks for your question. If you want to use Poly with your dataset. You can refer to the unit test:

https://github.com/microsoft/mttl/blob/main/tests/test_routed_multi_expert_model.py#L257

If you are interested in reproducing the results from the paper.

You need to use the branch and follow the instructions:

git checkout mhr_camera_ready

[kadarsh22]

Thank you @shuishen112 for your response, and apologies for the lack of clarity in my earlier comment. I was referring to Table 1 in the paper "Towards Modular LLMs by Building and Reusing a Library of LoRA", where results are reported on the Phi-2 model using the Polytropon and MHR methods.

I'm currently trying to reproduce those baseline results, specifically the performance of Phi-2 on the Flan dataset using Polytropon and MHR. Could you please guide me on how to replicate this setup?

Thank you!

[kadarsh22]

I tried following [https://github.com/microsoft/mttl/blob/main/tests/test_routed_multi_expert_model.py#L257] as you suggested. I had a small confusion here, model modifier is to be set to "skilled_lora" for MHR, while for poly it is to be set as 'lora' right?