How to reproduce the OLMo 3 SFT stage and obtain OLMo-3-7B-Instruct?

Hi, thanks for releasing the training code!

I have a few questions about reproducing the SFT stage of OLMo 3.

If I directly run the SFT script provided in this repository, will it reproduce the OLMo-3-7B-Instruct model?

In particular, should BASE_CKPT="/weka/oe-training-default/ai2-llm/path/to-base/checkpoint/step12345/model_and_optim" point to the checkpoint of OLMo-3-1025-7B?

According to the model-flow figure in the OLMo 3 paper, there appears to be a separate Think SFT stage before Instruct SFT. Therefore, I am not sure whether the released SFT script alone is sufficient to reproduce OLMo-3-7B-Instruct, or whether an additional Think SFT stage is required first.

Could you please provide more details about the exact SFT pipeline used in the paper, including the starting checkpoint and the sequence of SFT stages needed to reproduce the released model?

Thanks!

Start with the released SFT script and its BASE_CKPT setting, then compare the script's flow with the model-flow figure in the OLMo 3 paper. Document the starting checkpoint and the ordered SFT stages required to reproduce OLMo-3-7B-Instruct.

由索引模型根據 Issue 內容生成。