Fine-Tune LLM for Smart Assistant
Publicada el 2026-07-14
Descripción de la oferta
I am developing a conversational home-assistant that can • control lights, switches, thermostats, climate systems, security cameras and other Home Assistant-compatible devices, • deliver real-time weather and news briefings, and • manage household schedules and reminders. To reach that goal I need an AI/ML engineer who has hands-on experience fine-tuning Large Language Models such as Gemma, Llama-2/3, Qwen or Mistral. You should be comfortable running LoRA or QLoRA pipelines on Hugging Face Transformers and PyTorch, preparing conversational JSONL datasets, and iterating quickly on prompt- and instruction-tuning strategies. Beyond raw fine-tuning, the model must: • handle tool / function calling so it can trigger Home Assistant APIs, • return structured outputs that downstream automations can parse, and • be delivered in a lightweight, quantised format (INT4 or GGUF) for on-prem GPU and potential edge deployment. You’ll also be responsible for evaluation, optimization and, when needed, multilingual extensions. Familiarity with MCP-style tool-calling frameworks, Linux, Git and GPU training workflows is expected. Deliverables 1. Curated, reproducible training codebase (Git) with data-cleaning scripts. 2. Fine-tuned checkpoint plus quantised variants. 3. Sample inference script showing successful control of the three device classes above. 4. Short report covering metrics, prompt templates and recommended next steps. When applying, please link to previous LLM fine-tuning work, your Hugging Face or GitHub profile, and any deployed models or demos that showcase comparable functionality.
Skills
Fuente original: freelancer