yulan-hybrid-cpt / README.md
O2iginal's picture
Add files using upload-large-folder tool
5d791cc verified
|
Raw History Blame Contribute Delete
831 Bytes

YuLan hybrid continued-pretraining checkpoints

Archive upload is in progress. Do not assume a checkpoint is complete until UPLOAD_COMPLETE.json exists.

  • Stage15 final / Stage16a starting checkpoint: iter_0006855.
  • Stage16a final checkpoint: iter_0002384.
  • checkpoints/: native Megatron Core distributed checkpoints, including optimizer/training state as present in the source.
  • hf/: the existing Hugging Face exports recovered from the YuLan object store.
  • Training logs and launch scripts retain the stage-relative layout. HF export logs are under hf_export_logs/.
  • Files listed in archive-manifest.json are verified against remote sizes and content hashes at completion.

This repository is public at the owner's request. Original training checkpoint files are read in place through task-owned hardlinks.