4cee/raze-v2-gemma3n-e4b

This is a custom QLoRA fine-tune of Gemma-3n-E4B-it. It's trained on online conversations of my own friend group, with consent.

Disclaimer: this model is STILL VERY UNSTABLE. Most times it generates half-legible nonsense. Be weary!

On a related note; it will just hallucinate usernames. Or respond as multiple users.

This is the second model I've trained on this dataset. Due to the unfortunate nature of the dataset, it's still weird and stupid. If you want a more stable model still with a distinct personality, use raze-v3-hybrid, or raze-v3-calcium. Unfortunately it does not have the visual capabilities of the base model. I don't know how to keep them and it would require a lot of difficulty trying to make it like that.

Half of the training data was formatted as such:

{"messages": [{"role": "user", "content": ""Below is a chat log. Continue the conversation as [username1]. \n\n[username1]: [message1]\n[username2]:[message2]\n\n"}, {"role": "assistant", "content": "[response message]"}]}

And the other half was formatted like this:

{"messages": [{"role": "user", "content": "Roleplay as [username]. Reply to the following message.\n\n[message]\n\n"}, {"role": "assistant", "content": "[response]"}]}

This way, it can handle both one-on-one conversation, and conversation as a group. If you want the most accurate responses (why would you, it's funnier without it), then use something like that. I think it's best suited as an automated application.

Not much is actually different in terms of the model itself compared to v1. However I think the splitting of the dataset helped to mellow it out a little bit. It should be more legible.

Sorry if this doesn't fit whatever huggingface standards stuff. I see a lot of models split into the 4 different safetensors files and I deleted those. You get ggufs, deal with it i guess???

Disclaimer 2: I have no idea how to use HuggingFace, Github, or literally anything like that. This was a minor project that I did for fun. All of this was vibe coded with the help of Gemini-3-preview. Please forgive me if anything goes wrong.

Disclaimer 3: This model was not trained for the explicit purpose of generating anything harmful or against the Gemma Prohibited Use Policy. I did my best to filter out such content, however it may still be present and/or randomly generated by the model. Please don't sue me Google, oh god.

License and Terms

This model is a derivative of Gemma 3n E4B by Google.

Gemma is provided under and subject to the Gemma Terms of Use found at https://ai.google.dev/gemma/terms.

By using this model, you agree to the Gemma Terms of Use and the Prohibited Use Policy. (https://ai.google.dev/gemma/prohibited_use_policy)