Fine-tuning code is not available in repo

[shubhamguptalnmiit]

Hello, thank you for this excellent work. I was hoping to reproduce it but couldn't find the fine-tuning code file. Please guide.

[dreasysnail]

Hi @shubhamguptalnmiit we used Nvidia Megatron (https://github.com/nvidia/megatron-lm) for the training. It would still need time to work on a solution with Apple internal team to release the code interleaved with 3rd party code, as it involves some copyright issues. The training is quite straightforward as it's using standard SFT recipe on the generated data, and it might not necessarily use the exactly same training infrastructure as what we used. We expect it might be easier for you to directly use your favorite SFT pipeline (e.g. Megatron, Llama factory, Torchtune, etc...) while we figure out a clean version of our code for release.

We target to release the data generation code in two weeks. Please stay tuned. Thanks!

[richardbaihe]

Hi @shubhamguptalnmiit , thanks for your interests. We just released the data generation pipeline here: 15c4292, which outputs the data in the format of Megatron-Bridge training pipeline. Hope this helps and please feel free to reach out if you have any further questions!