colawolfie/Chen-9B

Overview

Chen-9B is a 9 billion parameter model, fine-tuned from Qwen/Qwen3.5-9B-Base, with the primary goal of emulating the distinctive chat style of the older "c.ai 1.1" model from Character.AI. This was achieved through distillation from Character.AI outputs, leveraging the PIPPA dataset and additional private datasets.

Key Characteristics

• Character Emulation: Successfully replicates the conversational quirks and style of the original Character.AI 1.1, including some of its known eccentricities.
• Roleplay Focus: Optimized for interactive roleplay and creative writing scenarios, expecting character descriptions in the system prompt.
• Training: Trained in 4-bit QLoRA mode and merged into a full bfloat16 model for ease of deployment.
• No Safety Alignment: Lacks built-in content filters or safety alignment, meaning it can generate NSFW or offensive content.
• Not for Factual Information: Not designed for truthfulness, high-quality coding, or instruct tasks; should not be used as a source of factual information.

Usage Recommendations

• Chat Template: Uses the standard ChatML format.
• Sampling Parameters: Recommended temp between 0.8-1.2, top_k 40, top_p 0.95, min_p 0.05.
• Repetition Penalty: Use rep_pen between 1.01-1.09 to mitigate repetition, a common trait inherited from its source data.
• System Prompt: Best used with a character description in the system prompt for intended roleplay behavior. Without it, it may act as an AI assistant or a random character.

Licensing

• Licensed under the Apache License 2.0, similar to its base model, Qwen3.5-9B-Base.