Delta-Vector/Rei-24B-KTO
Delta-Vector/Rei-24B-KTO is a 24 billion parameter language model with a 32768 token context length, developed by Delta-Vector. This model is specifically fine-tuned for creative writing and roleplaying, aiming to replicate the prose style of Anthropic Claude models. It utilizes a two-step training process involving Supervised Fine-Tuning (SFT) on the PaintedFantasy dataset and subsequent KTO (Kahneman-Tversky Optimization) to enhance coherency and instruction following.
Loading preview...
Rei-KTO-24B: Creative Prose and Roleplaying Model
Delta-Vector's Rei-KTO-24B is a 24 billion parameter language model designed for creative writing and roleplaying applications. It aims to emulate the prose style and coherency found in Anthropic's Claude models (Opus and Sonnet).
Key Capabilities & Features
- Creative Writing & Roleplaying: Optimized for generating engaging narratives and character interactions.
- KTO Enhanced: Utilizes Kahneman-Tversky Optimization for improved coherency and instruction adherence.
- Claude-like Prose: Fine-tuned to replicate the sophisticated writing style of Anthropic Claude models.
- V7 Tekken Prompting: Supports a specific prompting format for consistent interaction.
Training Details
The model underwent a two-step training process:
- Supervised Fine-Tuning (SFT): Initial training on the Zerofata's PaintedFantasy dataset, noted for its effectiveness with anime-otaku-esque characters.
- KTO Optimization: Subsequent training to further enhance coherency and instruction following capabilities.
Training involved 2 epochs over 24 hours using 8 x A100 GPUs, built with Axolotl.
Available Formats
Quantized versions are available for various deployment needs:
- GGUF Format for LLama.cpp & forks.
- EXL3 Format for TabbyAPI.