ConicCat/humans.txt-Diverse-WPO-24B
ConicCat/humans.txt-Diverse-WPO-24B is a 24 billion parameter model post-trained from Mistral Small 3 Base, specifically designed for creative writing and roleplay. It emphasizes maintaining output diversity and creativity without compromising quality, utilizing an adapted diverse WPO implementation and exclusively human-generated data. This model excels in generating varied and high-quality creative text, making it suitable for applications requiring imaginative and diverse narrative outputs.
Loading preview...
Model Overview
ConicCat/humans.txt-Diverse-WPO-24B is a 24 billion parameter language model built upon the Mistral Small 3 Base architecture. Its core focus is on creative writing and roleplay, distinguishing itself through a unique approach to maintaining output diversity and creativity during the alignment process.
Key Differentiators
- Diverse WPO Implementation: The model incorporates a diverse WPO (Weighted Policy Optimization) adaptation, inspired by research on modifying LLM post-training for creative writing, ensuring varied and imaginative outputs.
- Human-Only Data: It is post-trained exclusively on human-generated data, avoiding synthetic data to preserve natural language patterns and creative nuances.
- Quality Preservation: Despite its emphasis on diversity, the model is designed to maintain high output quality throughout its alignment.
Recommended Usage
For optimal performance, particularly given its Mistral Small 3 Base heritage, specific sampler settings are recommended:
- Chat Template: Alpaca
- Temperature: 0.6
- Top P: 0.95
- Repetition Penalty: 1.08
- Repetition Range: 4096
Ideal Use Cases
This model is particularly well-suited for applications requiring:
- Generating diverse and imaginative creative narratives.
- Engaging in nuanced and varied roleplay scenarios.
- Tasks where maintaining output creativity without sacrificing quality is paramount.