weepiess2383/arc-gemma4b-bb75k

Archive: gemma4b-bb75k

EMA-consolidated research checkpoint archive.

Gemma-3-4B backbone fair pretrain (bb75k schedule). Backbone-comparison line.

Each *.shards.stepNNNNNN/state.pt holds only the EMA payload (state['extra']['ema'], fp32, clean names) extracted from the original training checkpoint, fingerprint-verified against engine_meta.pt. Loadable directly by methods/lafm/vla_flow_ft.py:load_pretrain_trainables with init_use_ema=true. Optimizer states were not retained.

Manifest

[
  {
    "run": "lafm_vla_flow_gemma4b_gemma4b_bb75k",
    "step_dir": "lafm_vla_flow_gemma4b_weights.ckpt.shards.step000075000",
    "source": "ema",
    "bytes": 2152268130
  }
]