Hugging Face's logo

tapiocaTakeshi
/
Qubit-7B-ja-training

Model card Files
xet
 Community
Qubit-7B-ja-training / code

Commit History

  1. train: per-rank FSDP flat-param + optimizer shard checkpoints (replace DCP), local-first then bucket copy

    tapiocaTakeshi
    Verified signature
    518d1d7

  2. train: DCP to local disk + sequential copy to bucket; resume copies bucket->local before load

    tapiocaTakeshi
    Verified signature
    be94448

  3. train: flatten optimizer state dict for DCP (fix multi-rank restore)

    tapiocaTakeshi
    Verified signature
    af53195

  4. train: FSDP scalar-param fix (lambda_base->[1]), direct safetensors write + SAVED_OK marker, 2h PG timeout

    tapiocaTakeshi
    Verified signature
    ce344e8

  5. prep: stream FineWeb2 test split (avoid preparing full train split)

    tapiocaTakeshi
    Verified signature
    51b49bd

  6. Add FSDP training script for Qubit 7B (QBNN)

    tapiocaTakeshi
    Verified signature
    65427b9

  7. Add data prep + tokenizer script

    tapiocaTakeshi
    Verified signature
    9730675