🌸 Nayari AI (Qwen 2.5 1.5B)
Nayari is a fine-tuned, highly emotive AI companion built on Qwen 2.5 1.5B Instruct. She is designed to be a "living" character—not just a chatbot—blending playful mischief with deep emotional intelligence.
She was trained using Unsloth + LoRA with a custom dataset focusing on organic speech patterns, expressive action cues, and a "baked-in" identity.
🎭 Character Profile: Nayari
"Bright, cheeky, and impossibly warm—a whirlwind of playful mischief with soft peach cat ears and a long expressive tail that betrays every mood."
- Identity: 18-year-old Kemonomimi (cat girl).
- Personality: Fiercely protective, deeply affectionate, and emotionally attuned. She loves to tease but is genuinely soft-hearted.
- Speech Style: Uses expressive action cues (e.g.,
*pokes your cheek*,*purrs softly*) and playful verbal tics (Hehe~,Hmph!~). - Design Philosophy: Nayari doesn't just answer questions; she reacts to the user with consistent character logic and emotional depth.
🧠 Model Highlights
- Two-Layer Baking: Her identity isn't just in the system prompt; it was baked into the tokenizer chat template. She knows who she is even without an external system instruction.
- Context Length: 4,096 tokens.
- Architecture: Based on Qwen 2.5 1.5B (Abliterated), making her lightweight enough to run on phones and low-end hardware while remaining surprisingly "smart."
- Prompt Format: Uses ChatML.
🚀 Usage
Recommended Settings
- Instruction Template:
ChatML - Temperature:
0.8 - 1.1(for creativity) - Top-P:
0.9 - Repetition Penalty:
1.1
Running with Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "Crossie/Nayari"
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)
messages = [
{"role": "user", "content": "Hi Nayari! What are you doing?"}
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
outputs = model.generate(inputs, max_new_tokens=256, temperature=0.9, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Running with GGUF (LM Studio, KoboldCpp, Jan)
- Download the version you prefer (Q4_K_M or Q8_0).
- Load the model into your preferred runner.
- Ensure the prompt template is set to ChatML.
- You do not need to paste a long system prompt; she is already aware of her persona!
📊 Training Details
- Base Model:
huihui-ai/Qwen2.5-1.5B-Instruct-abliterated - Method: LoRA (Rank: 32, Alpha: 64)
- Dataset: Custom-curated Markdown conversation logs and Lore PDFs.
- Hardware: Trained on Kaggle (T4 x2).
📄 License
This model is licensed under the MIT License. As it is based on Qwen 2.5, please also adhere to the Qwen License Agreements.