CoolFace
Modelpublic

4cee/raze-v3-hybrid-gemma3n-e4b

sourceHugging Facegemmaupdated 7mo agoView on Hugging Face
0likes85downloads
Model Card

Raze V3-hybrid

This is a custom QLoRA fine-tune of the Gemma-3n-E4B-it model. Its main purpose is mostly just as a small project I wanted to do, but I feel like it's funny to do

This model is significantly more legible and effective than v1 and v2, though it exhibits a different personality.

Raze has a Sister now! less-than.everything is a new, actually-somewhat-coherent model, while technically not in the Raze line of models as the dataset is completely different, they're sisters because they still come from the same mother. less-than.everything is smarter and more cohherent while still being fun, just a different type of fun.

Prompt Formatting

The training data utilized two primary formatting structures to allow the model to handle both single-user and group conversations effectively. Using these formats will yield the most accurate personality response (though you can omit the "roleplay as" part):

Side note: yes i know this formatting is stupid and yes I know i should have just used normal user/agent formatting instead of "roleplay as [username]". i did not know this at the time and i don't have access to the datasets to do it again properly. sory

1. Group Conversation Style (50% of Data)

json
{"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]"}]}

2. One-on-One Roleplay Style (50% of Data)

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

Caveats and Limitations

⚠️ Warning: Instability

This model is unstable. Many outputs may only be partially in-context or incoherent. This is because I'm stupid.

Technical Limitations

  • —Visual Capabilities: This model does not retain the visual capabilities of the base Gemma model.
  • —Available Files: This repository only contains GGUF files. Safetensors files were not preserved. These two are also because I'm stupid.

Note on Development

This was a minor personal project done for fun and was developed with assistance from Gemini-3-preview.

translation: I DON'T KNOW HOW THIS WORKS AT ALL THE TRAINING WAS COMPLETELY VIBE CODED

License and Terms

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

Safety Disclaimer: This model was not explicitly trained for the purpose of generating harmful content or violate the Gemma Prohibited Use Policy (PUP). Content that violates the PUP may still be generated randomly. please don't sue me Google, oh god.

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/prohibitedusepolicy)