Overview
ConicCat/humans.txt-Diverse-OrPO-24B is a 24 billion parameter model built upon the Mistral Small 3 Base architecture. Its core focus is on creative writing and roleplay, specifically engineered to deliver diverse and imaginative outputs while preserving high quality. The model's unique alignment process involves post-training exclusively on human-generated data and leveraging the Diverse OrPO method, as detailed in the paper "Modifying Large Language Model Post-Training for Diverse Creative Writing" (arXiv:2503.17126).
Key Capabilities
- Enhanced Creative Writing: Generates varied and high-quality text for creative applications.
- Roleplay Optimization: Designed to excel in interactive roleplaying scenarios.
- Diversity Maintenance: Aligned to prevent output homogenization, ensuring a wide range of creative responses.
- Human Data Training: Benefits from post-training solely on human-sourced data, avoiding synthetic content.
Recommended Usage
For optimal performance, particularly given its sensitivity to sampler settings, the model recommends specific configurations:
- Chat Template: Alpaca
- Temperature: 0.6
- Top P: 0.95
- Repetition Penalty: 1.08
- Repetition Range: 4096
This model is ideal for developers and creators seeking an LLM that prioritizes imaginative and diverse textual generation for narrative, interactive, and creative content applications.