|
|
--- |
|
|
base_model: |
|
|
- allura-org/Gemma-3-Glitter-27B |
|
|
base_model_relation: quantized |
|
|
library_name: transformers |
|
|
tags: |
|
|
- mergekit |
|
|
- merge |
|
|
--- |
|
|
## Quantized using the default exllamav3 (0.0.3) quantization process. |
|
|
|
|
|
- Original model: https://huggingface.co/allura-org/Gemma-3-Glitter-27B |
|
|
- exllamav3: https://github.com/turboderp-org/exllamav3 |
|
|
--- |
|
|
# ✨G3 Glitter 27B✨ |
|
|
<figure> |
|
|
<img src="https://huggingface.co/ToastyPigeon/Gemma-3-Glitter-27B/resolve/main/ComfyUI_02512_.png" width="600"> |
|
|
</figure> |
|
|
|
|
|
A creative writing model based on Gemma 3 27B. |
|
|
|
|
|
[Columbidae/gemma-3-27b-half](https://huggingface.co/Columbidae/gemma-3-27b-half), a 50/50 merge of 27B IT and 27B PT, was used as the base model. (This was done because of the success of [Starshine](https://huggingface.co/ToastyPigeon/Gemma-3-Starshine-12B), a 50/50 IT and PT merge.) |
|
|
|
|
|
The inclusion of PT model does weaken the instruct, but it also weakens the censorship/hesitancy to participate in certain fictional stories. The prose also becomes more natural with less of the IT model included. |
|
|
|
|
|
**This model does better with short and to-the-point prompts. Long, detailed system prompts will often confuse it.** (Tested with 1000-2000 token system prompts to lackluster results compared to 100-500 token prompts). |
|
|
|
|
|
## Instruct Format |
|
|
|
|
|
Uses Gemma2/3 instruct and context. Like Glitter 12b, this works well with `temp = 1, top-nsigma = 1.5`. |
|
|
``` |
|
|
<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> |
|
|
``` |
