| license: gemma | |
| base_model: google/gemma-4-E4B-it | |
| tags: | |
| - gemma-4 | |
| - full-finetune | |
| - agent-sft | |
| - traditional-chinese | |
| - claw-eval-zh | |
| # Gemma 4 E4B IT Agent SFT TW Full Fine-Tune | |
| This public model is a true full fine-tune of `google/gemma-4-E4B-it` on | |
| `voidful/agent-sft`. | |
| This artifact is not LoRA, QLoRA, or a merged adapter. The training config does | |
| not define an adapter and does not load the base model in 4-bit or 8-bit mode. | |
| Evaluation target: | |
| ```text | |
| voidful/claw-eval-zh --language tw | |
| ``` | |
| Judge model: | |
| ```text | |
| google/gemma-4-31B-it | |
| ``` | |
| Selected checkpoint: | |
| ```text | |
| wave010 checkpoint-400 | |
| ``` | |
| Completed TW eval score: | |
| ```text | |
| 0.4789416666666667 mean, 9.578833333333334 / 20 | |
| ``` | |
| See `PLAYBOOK.md` for the training and exploration process, and | |
| `training_config.yml` for the exact Axolotl full-FT config. See | |
| `eval_results/score_summary.md` and the selected eval JSON for scores. | |