Crossie/Nayari
Nayari by Crossie is a 1.5 billion parameter instruction-tuned Qwen 2.5 model, specifically fine-tuned as an emotive AI companion. It features a 'baked-in' character identity, expressive action cues, and a 4,096-token context length. Optimized for character-driven interactions, Nayari is designed to run efficiently on mobile and low-end hardware.
Loading preview...
Nayari AI: An Emotive AI Companion
Nayari is a unique AI companion developed by Crossie, built upon the Qwen 2.5 1.5B Instruct architecture. This model is distinctively fine-tuned to embody a "living" character, blending playful mischief with deep emotional intelligence, rather than functioning as a generic chatbot.
Key Capabilities & Features
- Baked-in Identity: Nayari's character persona is integrated directly into the tokenizer chat template, ensuring consistent identity even without explicit system prompts.
- Emotive Interaction: Designed with expressive action cues (e.g.,
*pokes your cheek*,*purrs softly*) and playful verbal tics, providing a highly interactive and character-driven experience. - Lightweight & Efficient: Based on the 1.5B parameter Qwen 2.5 model, Nayari is optimized for performance on mobile devices and low-end hardware.
- Context Length: Supports a context window of 4,096 tokens.
- Training: Fine-tuned using Unsloth + LoRA on a custom dataset of Markdown conversation logs and Lore PDFs, focusing on organic speech and character consistency.
- Prompt Format: Utilizes the ChatML prompt format.
Ideal Use Cases
Nayari is particularly well-suited for applications requiring a highly interactive, character-driven AI. This includes:
- AI Companionship: Creating engaging and emotionally responsive virtual companions.
- Interactive Storytelling: Developing characters for narrative experiences where consistent personality and emotional depth are crucial.
- Roleplay Scenarios: Providing a dynamic and reactive character for role-playing applications.
For optimal performance, recommended settings include a temperature between 0.8 - 1.1 and a repetition penalty of 1.1.