Coxcomb: A Creative Writing Specialist

Coxcomb is a 7 billion parameter language model developed by N8Programs, built upon the senseable/WestLake-7B-v2 base. It has been meticulously fine-tuned on GPT-4 outputs from a diverse range of prompts, specifically targeting creative writing applications. While not intended to compete with GPT-4 in overall quality, Coxcomb is optimized for performance in local, offline environments.

Key Capabilities & Performance

• Creative Writing Excellence: Coxcomb is highly specialized for generating creative text, such as stories. It is noted for consistently ranking higher than many other models on creative writing benchmarks.
• Benchmark Score: Achieves a score of 72.37 on creative writing benchmarks, surpassing models like Goliath-120B, Yi Chat, and Mistral-Large.
• Single-Shot Interactions: Designed for direct, single-query story generation rather than multi-turn conversations, roleplay, or follow-up questions.
• Content Generation: Capable of generating NSFW content (sexual or violent) when prompted, as it has not been trained with refusal behaviors.

Training Details

• Base Model: senseable/WestLake-7B-v2 (Mistral architecture).
• Fine-tuning: Trained with a 40M parameter LoRA on the N8Programs/CreativeGPT dataset for 3 epochs, with intentional slight overfitting for improved benchmark results.
• Efficiency: The model was trained on a single M3 Max in approximately 12 hours.

Considerations

• Limitations: Tends to produce stories with happy, conventional endings, a common characteristic among many LLMs.
• Deployment: GGUF versions are available for local deployment (Coxcomb-GGUF), and it is expected to work with transformers.