Model Overview
Delta-Vector's François-PE-V2-Huali-12B is a 12 billion parameter language model designed to produce creative and refreshing prose. This model is a direct evolution of the original François-PE/Huali series, building upon the Dans-PE-12B base model which was fine-tuned using a diverse dataset including light novels, books, and roleplay logs.
Key Enhancements & Features
- KTO Enhancement: Utilizes KTO (Kahneman-Tversky Optimization) to significantly increase text coherency and refine prose quality.
- Unique Writing Style: Specifically trained to offer a distinct writing style, differentiating it from other NeMo-based models, focusing on short and sweet narrative outputs.
- ChatML Formatting: Optimized for ChatML prompting, ensuring straightforward integration into conversational AI applications.
- Extended Context Length: Supports a context window of 32768 tokens, allowing for more extensive and detailed interactions.
Training Details
The model underwent a single epoch of fine-tuning on 8 x H200 GPUs, with resources provided by Kalomaze. The training process leveraged the Axolotl framework.
Recommended Use Cases
- Creative Writing: Ideal for generating imaginative stories, character dialogues, and descriptive passages.
- Roleplay Scenarios: Excels in creating engaging and coherent responses for roleplaying applications.
- Prose Generation: Suitable for tasks requiring a unique and refined textual output style.
Users are recommended to employ either Euryale's or EVA's system prompts for optimal performance, with an example of Sao10k's Euryale System Prompt provided for guidance.