Homie -- Personal AI Assistant

Homie is a privacy-first, local AI assistant with identity baked into the weights. No system prompt needed. The model inherently knows it's Homie.

Features

• Identity: Knows it's Homie, not a generic chatbot
• Tool-aware: Understands email, calendar, files, git, web search, voice, mesh
• Privacy-first: Emphasizes local-only, no cloud
• Concise: Leads with answers, code-first for dev questions
• Secure: Refuses prompt injection, never leaks credentials
• Multi-device: Understands mesh networking and device sync
• Learning: Remembers preferences across sessions

Training

• Base model: LiquidAI/LFM2.5-1.2B-Instruct (Apache 2.0)
• Method: QLoRA (rank 64, alpha 128, 0.54% trainable)
• Data: 69 hand-crafted examples across 10 categories (x3 = 207 total)
• DPO: 10 preference pairs (Homie-style vs generic)
• Categories: identity, tools, mesh, memory, security, coding, proactive, errors, flow, system
• Epochs: 5, Loss: 1.414, Accuracy: 92%

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Muthu88/Homie", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("Muthu88/Homie", trust_remote_code=True)

messages = [{"role": "user", "content": "Who are you?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Part of the Homie AI Project

• GitHub: https://github.com/MSG-88/Homie
• Website: https://heyhomie.app