Another one of those test models

we were chatting and just decided: "What would happen if we have a higher learning rate and run 3 epochs🤓"

So here it is, the second generation of filtered Gemma3NPC, with bare minimum effort of typing 6 characters and a little patience(3 hours).

Again, our training notebook is on Github.

This time, we accidentally stopped the training when it reach step 200, so when we resumed, the training started from scratch but seems to have used the last checkpoint.

Here is a graph of the training loss, saved after after 5 steps.

Our top priority is now the gathering of more datasets, such as SODA and some real video game data.

We might try to switch to a new model(Qwen?), as the Gemma3n license is a little restrictive.

New methods other than SFT to improve performance, like GRPO.