AliceThirty/G4-Alice-v1.2-31B
AliceThirty/G4-Alice-v1.2-31B is a 31 billion parameter roleplay fine-tune of the Serenity model, developed by AliceThirty. This version specifically restores and enhances reasoning capabilities by training the model to utilize thinking blocks. It is optimized for creative and uninhibited roleplay scenarios, offering improved long-term memory and character consistency compared to its predecessors. The model is designed for engaging interactive storytelling and character-driven applications.
Loading preview...
Overview
AliceThirty/G4-Alice-v1.2-31B is a 31 billion parameter language model specifically fine-tuned for roleplay scenarios. It is based on the Serenity-31B model and utilizes the same dataset as MeroMero-31B. This version, v1.2, significantly restores and enhances reasoning capabilities by training the model to consistently open and use 'thinking blocks' during generation, leading to more coherent and intelligent responses.
Key Improvements & Features
- Enhanced Reasoning: Version 1.2 focuses on integrating reasoning, which was a bug in earlier versions, by training the model to use internal thought processes.
- Optimized for Roleplay: Fine-tuned on a dataset geared towards creative and uninhibited roleplay, aiming for good long-term memory, character consistency, and engaging writing.
- Low Learning Rate Fine-tuning: Trained with a low learning rate (7e-6) and a LoRA rank of 80 for one epoch, ensuring the fine-tune complements rather than overrides the base Serenity model's characteristics.
- Performance: The fine-tuning process was efficiently conducted on a single RTX 5090 using Unsloth.
Recommended Settings
For optimal performance, the following sampling parameters are suggested:
- Temperature: 0.8 - 1.2
- Min_p: 0.05
- Skip Special Tokens: false
Users seeking faster replies without the explicit reasoning blocks might prefer version 1.1.