We trained this model as a rank-12 LoRA adapter with one epoch over pippa_filtered 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 a 150-step warmup. With a gradient clipping of 0.5.

Check out our training notebook here.

Here is a graph of the Step Training Loss, saved every 5 steps:

chimbiwide/Gemma3NPC-filtered

Gemma3NPC-filtered-float16

The "filtered" model that delivers censored general role-playing at great speed.

chimbiwide/
Gemma3NPC-filtered