About the weight provided in this release after which training stage

[DFLyan]

The training in the paper is roughly divided into four parts, as shown in the following figure：

What I want to know is which training stages the weights provided on the Hugging Face have gone through. For example, is it after action injection training or after SFT.

[DFLyan]

and if I want to use another vlm, like qwen3vl-2b, as my backbone to train a model from scratch, based on the released code, which configurations should I modify.
Thank you!

[super-anova]

Hi, thanks for your question.

The first question in the Alpamayo 1.5 FAQ section (https://github.com/NVlabs/alpamayo1.5#frequently-asked-questions-faq) "How does Alpamayo 1.5 relate to Alpamayo 1?" may help you with your questions.

Alpamayo 1 was not RL post trained, while Alpamayo 1.5 was RL post trained.

[HuangChiEn]

Hi, thanks for your question.

The first question in the Alpamayo 1.5 FAQ section (https://github.com/NVlabs/alpamayo1.5#frequently-asked-questions-faq) "How does Alpamayo 1.5 relate to Alpamayo 1?" may help you with your questions.

Alpamayo 1 was not RL post trained, while Alpamayo 1.5 was RL post trained.

TKS for your explanation, however i also wonder that :
In FINETUNE_SFT.md, why we need to download the checkpoint of nvidia/Alpamayo-R1-10B before do the stage 1 VLM-SFT ? and trajectory tokenizer seems also depend on the checkpoint ?
Isn't stage 1 VLM-SFT should be trained on Cosmost-Reason (Qwen3VL-based) pretrained weight, instead of Alpamayo checkpoint, or nvidia/Alpamayo-R1-10B is cosmos-reason ?

https://github.com/NVlabs/alpamayo/blob/e6aefbc8d7a93480f045147bd36fcddeb9c829a2/docs/FINETUNE_SFT.md#download-the-checkpoint

[super-anova]

Hi @HuangChiEn. To clarify, the released finetuning scripts or recipes are intended for people to post train and develop further on the released Alpamayo models. So all the recipes start with downloading Alpamayo checkpoints. Btw, the recipes are now at https://github.com/NVlabs/alpamayo-recipes

This is separate from how we trained Alpamayo R1 10B from a pretrained Cosmos Reason 3 checkpoint.