The 16bit GGUF version of Gemma3NPC-Float16. We trained this model as a rank-16 LoRA adapter with one epoch over pippa using a 40GB vRAM A100 in Google Colab. For this run, we employed a learning rate of 2e-5 and a total batch size of 1 and gradient accumulation steps of 16. A cosine learning rate scheduler was used with an 800-step warmup. With a gradient clipping of 0.4. Check out our training notebook here.

Here is a graph of the Step Training Loss, saved every 10 steps: