Modelpublic

TitleOS/HomeGem4Bn

sourceHugging Facempl-2.0updated 11mo agoView on Hugging Face

0likes13downloads

Model Card

HomeGem4Bn

—Developed by: TitleOS
—License: mpl-2.0
—Finetuned from model : unsloth/gemma-3n-e4b-it-unsloth-bnb-4bit

HomeGem is a first generation finetune of Gemma 3n e4b IT, for use as a conversational LLM controlling your Home Assistant environment. Finetuning was performed on an A100 SXM4, and finetuned using the acon96/Home-Assistant-Requests dataset.

Eval Results

Step 5: Displaying post-training statistics and evaluation loss...
2797.18 seconds used for training.
Peak reserved memory: 17.799 GB.
Peak memory for training: 8.42 GB.

--- Evaluation Loss Progression ---
Step 20: Eval Loss = 2.0582
Step 40: Eval Loss = 1.6810
Step 60: Eval Loss = 1.5477
Step 80: Eval Loss = 1.4737
Step 100: Eval Loss = 1.4113
Step 120: Eval Loss = 1.3982
Step 140: Eval Loss = 1.4340
Step 160: Eval Loss = 1.4325
Step 180: Eval Loss = 1.4424
Step 200: Eval Loss = 1.4442

--- Summary ---
Initial Eval Loss: 2.0582
Final Eval Loss:   1.4442
Improvement:       0.6140
---------------------------------