Gemma 3 Starshine 12B Alt

🚀 🌠G3 Starshine 12B - Alt🌠

This project presents a creative writing model based on the merge of fine - tunes on Gemma 3 12B IT and Gemma 3 12B PT. It focuses on role - playing (RP) and offers better performance in turn - based chats.

This was Merge A2 in the testing set.

🚀 Quick Start

This is a creative writing model created by merging fine - tunes on Gemma 3 12B IT and Gemma 3 12B PT. It is the RP Focused merge, which can handle separate characters in turn - based chats with less impersonation. You can also check out the main Story Focused version.

This model is a merge of two G3 models:

• allura-org/Gemma-3-Glitter-12B - A merge of a storywriting and RP train (both also by ToastyPigeon), trained on instruct.
• ToastyPigeon/Gemma-3-Confetti-12B - An experimental application of the Glitter data using base instead of instruct, additionally including some adventure data in the form of SpringDragon.

The combination results in a great blend of Glitter's instruction - following ability and Confetti's free - spirit prose, reducing much of the hesitancy in Glitter.

Vision works if you pair a GGUF of this model with an appropriate mmproj file. The developer intends to fix the missing vision tower and make this model properly multimodal in the near future.

Thank you to jebcarter for the idea to make this. I love how it turned out!

📚 Documentation

Instruct Format

The model uses Gemma2/3 instruct and has been trained to recognize an optional system role.

⚠️ Important Note

While the model won't immediately reject the system role, the results may be better without it.

<start_of_turn>system
{optional system turn with prompt}<end_of_turn>
<start_of_turn>user
{User messages; can also put sysprompt here to use the built - in g3 training}<end_of_turn>
<start_of_turn>model
{model response}<end_of_turn>

Merge Configuration

A higher percentage of Glitter gives this model better turn - based instruct following, but it may be more rigid compared to the Story Focused version.

models:
  - model: ToastyPigeon/Gemma-3-Confetti-12B
    parameters:
      weight: 0.3
  - model: allura-org/Gemma-3-Glitter-12B
    parameters:
      weight: 0.7
merge_method: linear
tokenizer_source: allura-org/Gemma-3-Glitter-12B

Model Information