Overview

ChatWaifu_v1.4 is a 12 billion parameter CausalLM developed by spow12, specifically engineered to emulate visual novel characters. It is a merged model, built using mergekit from several base models including spow12/ChatWaifu_modify_data, anthracite-org/magnum-v2-12b, and mistralai/Mistral-Nemo-Instruct-2407, with NeverSleep/Lumimaid-v0.2-12B as its finetuned base. The model is primarily in Japanese and has undergone updates to refine its data format and apply filtering, as well as incorporating preference learning in its training pipeline.

Key Capabilities

• Visual Novel Character Emulation: Designed to act like characters from visual novels, supporting a range of personalities from popular titles like Senren＊Banka and Café Stella.
• Extended Context & Memory: Features a 32768 token context window and robust memory abilities, preventing conversational drift and repetition even in dialogues exceeding 20-30 turns.
• Zero-Shot Persona Generation: Capable of adopting character personas based solely on their descriptions without explicit fine-tuning for each character.
• Fluent Japanese Chat: Delivers high-quality, fluent conversational performance in Japanese.

Use Cases

• Interactive Storytelling: Ideal for applications requiring AI characters that can maintain consistent personas and memory over long interactions.
• Visual Novel Development: Can be integrated into visual novel projects to power character dialogue and interactions.
• Research in Character AI: Useful for exploring character-driven AI and long-context conversational models.

Limitations

• The model was trained on Japanese datasets, including visual novel content that may contain NSFW material, and thus has the potential to generate NSFW content.
• Currently available for non-commercial and research purposes only.