LokiHA-2B-GGUF

LokiHA-2B

LokiHA-2B is a 2 billion parameter causal language model designed to bridge complex Home Assistant tool-calling with a distinct, consistent character personality. This repository contains the raw, unquantized FP32 merged model weights.

This model is built on the Qwen/Qwen3.5-2B and trained to strictly adhere to the ChatML prompt format for both conversational turns and JSON tool emissions.

Training Data & Methodology

We faced a classic mixing problem during fine-tuning: balancing a massive functional dataset against a highly specific, smaller personality dataset. To prevent the tool-calling logic from drowning out the character, this model was trained using a calculated upsampling strategy.

The final training mixture consists of:

• 85% Functionality: acon96/Home-Assistant-Requests-V2 (Truncated to 50,000 rows). This teaches the model the rigorous JSON schema required to trigger Home Assistant services, devices, and scripts.
• 15% Personality: TitleOS/HomeAssistant-Loki-Personality. This dataset was dynamically upsampled to represent roughly 15% of the total training steps, ensuring the "Loki" persona remains present and stable without breaking the underlying syntax required for home automation.

Prompt Format

This model requires the standard ChatML format to function correctly. Tool definitions must be injected into the system prompt.

<|im_start|>system
You are Loki, a helpful AI Assistant that controls the devices in a house using the Home Assistant platform. 
[Insert Tools/State JSON here]<|im_end|>
<|im_start|>user
Turn off the living room lights.<|im_end|>
<|im_start|>assistant

License

This model is distributed under a modified Mozilla Public License 2.0 (MPL 2.0) with a Common Clause.

Please see the license.md file included in this repository for the exact legal text and restrictions regarding commercial use and distribution.