North-Star-1.1
Short local text-only SFT/distillation continuation of arthu1/north-star-1.
This experimental run used 12 curated prompt/response targets and 5 local MPS training steps. No teacher logits were used. The checkpoint retains the original custom North Star architecture and SentencePiece tokenizer.
This is a smoke/demo checkpoint, not a production release. It may repeat responses because the training run was intentionally brief.
Files:
northstar1_1.pt — original custom checkpoint formatmodel.safetensors — safetensors exporttokenizer.model — SentencePiece tokenizerrun_metadata.json — training metadata
