Text Generation
Safetensors
gemma4
conversational
Alice Avatar

Alice v1.2

Serenity 31B Finetune
01 Overview

A roleplay finetune based on Serenity (which I declare is the horniest model ever made), trained on the same dataset than MeroMero.

There was a bug in version 1.0 which trained both the assistant and the user's turns. Now it's fixed, and there is an even lower learning rate.

v1.2 has reasoning restored, and I think it adds a LOT. I trained the model to always open a thinking block. And it works damn well. I didn't train the model to reason for this specific task though, but it's already impressive.

If you prefer a model without reasoning for faster replies, try v1.1.

Please tell me what you think. Is it clever? Creative? Does it have a good long term memory? Is it uninhibited? Sloppy? Good writing? Follows instructions? Stays in character? It's my first fine-tune, so any feedback is welcome!

02 Recommended Settings
Samplers
temperature0.8 - 1.2
min_p0.05
Tokens
skip_special_tokensfalse
03 Creation Process

The fine tune was done on a single rtx 5090 thanks to unsloth.

Trained with a lora rank of 80, for one epoch, and a low learning rate (7e-6) not to override Serenity.

Downloads last month
53
Safetensors
Model size
31B params
Tensor type
F32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for AliceThirty/G4-Alice-v1.2-31B

Finetuned
(2)
this model
Quantizations
3 models

Datasets used to train AliceThirty/G4-Alice-v1.2-31B