LoRA adapters (rank 16, alpha 32) from self-distilled supervised fine-tuning (SD-SFT) of Gemma 4 E4B on Koka, run 1-2-0.
Training used the 4-bit base unsloth/gemma-4-e4b-it-unsloth-bnb-4bit.
| Subfolder | Contents |
|---|---|
checkpoint-400 |
adapter, tokenizer and trainer state (optimizer, scheduler, random-number states) |
checkpoint-500 |
adapter, tokenizer and trainer state (optimizer, scheduler, random-number states) |
checkpoint-511 |
adapter, tokenizer and trainer state (optimizer, scheduler, random-number states); last saved checkpoint |
Load one checkpoint with PEFT, passing subfolder="checkpoint-511".
checkpoint-400/adapter_model.safetensors 170 MBcheckpoint-500/adapter_model.safetensors 170 MBcheckpoint-511/adapter_model.safetensors 170 MBcheckpoint-400/optimizer.pt 71 MBcheckpoint-500/optimizer.pt 71 MBcheckpoint-511/optimizer.pt 71 MBcheckpoint-400/rng_state.pth 14.6 kBcheckpoint-500/rng_state.pth 14.6 kBcheckpoint-511/rng_state.pth 14.6 kBcheckpoint-400/training_args.bin 5.8 kBcheckpoint-500/training_args.bin 5.8 kBcheckpoint-511/training_args.bin 5.8 kBcheckpoint-400/scheduler.pt 1.5 kBcheckpoint-500/scheduler.pt 1.5 kBcheckpoint-511/scheduler.pt 1.5 kBunsloth/gemma-4-E4B-it