Views
No views yet
checkpoint-315 = epoch 1, checkpoint-630 = epoch 2, checkpoint-945 = epoch 3. Each contains model.safetensors + DeepSpeed ZeRO shards.model.safetensors per checkpoint; global_step*/ folders hold DeepSpeed ZeRO optimizer shards (only needed to resume training). Use zero_to_fp32.py to reconstruct fp32 weights from shards if needed.