Overview

Delta-Vector/MS3.2-Austral-Winton is a 24 billion parameter model, a specialized fine-tune of Codex 24B, developed by Delta-Vector. This model is engineered to be a generalist roleplay and adventure model, focusing on enhancing narrative flow and character interaction.

Key Capabilities & Features

• Codex Finetune: Built upon the robust Codex 24B architecture.
• KTO Enhanced: Utilizes KTO (Kahneman-Tversky Optimization) alignment to refine coherency and improve overall writing quality.
• Multi-stage Fine-tuning: Underwent several stages of fine-tuning, including 4 epochs of SFT with a datamix similar to Francois-Huali/Austral 70B, followed by KTO, and then another 4 epochs with Rep_remover to eliminate repetitive outputs.
• Adventure/Roleplay Generalist: Specifically designed to excel in generating dynamic plots and engaging character dialogues for adventure cards and roleplay scenarios.

Training Details

The model's training involved 4 epochs of base SFT, followed by 1 epoch of KTO for coherency, and a final 4 epochs using Rep_remover. This process, totaling approximately 80 hours, was conducted on 8 x A100 GPUs.

Chat Format

This model uses the ChatML format for interactions, ensuring compatibility with common inference setups.

Use Cases

This model is particularly well-suited for applications requiring:

• Interactive Storytelling: Generating engaging and coherent narratives for adventure games or interactive fiction.
• Roleplay Scenarios: Creating dynamic and responsive character interactions.
• Creative Writing: Assisting in plot development and dialogue generation where high-quality, non-repetitive output is crucial.