Request for Nemotron 3 Intermediate Checkpoints for Safety Research

[tigist-far]

Dear Nemotron team,

I am a member technical staff at FAR.AI where we'd like to replicate parts of the Nemotron 3 training recipe with safety-enhancing modifications. As our interventions can happen at later stages of training, we can initialize with intermediate checkpoints to save on compute.

Thank you for having already released a broad set of models on HuggingFace as part of the NVIDIA Nemotron v3 collection. Would it be possible to release the intermediate checkpoints for Nemotron 3 as well to support our experiments?

In order of priority, we would be very grateful if you could release the following intermediate checkpoints as part of your HuggingFace collection:

• SFT Checkpoint for Nemotron-3-Super-120B-A12B (post-SFT, pre-RL)
• RLVR Checkpoint for Nemotron-3-Super-120B-A12B (post-RLVR, pre-SWE1)
• SWE1 Checkpoint for Nemotron-3-Super-120B-A12B (post-SWE1, pre-SWE2)
• SWE2 Checkpoint for Nemotron-3-Super-120B-A12B post-SWE2, pre-RLHF)
• SFT Checkpoint for Nemotron-3-Nano-30B-A3B (post-SFT, pre-RL)

@shashank3959 I am tagging you here considering you have triaged some of these requests in prior GitHub issues. We would be happy to provide further detail directly to you or the right team internally as appropriate. We look forward to your response.

Best regards,
Tigist