Further reading

This project

• Source code on GitHub — where to get it, and what these docs are built from
• The original codebase — the address the work was published under, and the one the citation names
• Model packages on Hugging Face — six exported packages, each shipping both an inference and a training stage
• Parameter-Efficient Tuning of Large Language Models on Mobile Devices — the master's thesis this framework accompanies
• SELLMA: Semantic Location through On-Device LLMs and WiFi Sensing — On-device semantic location recognition through privacy-preserving local LLMs with WiFi sensing.
• AI health agents on mobile, BMC Proceedings 2026, 20(12):A7 (EHRCON25 — openEHR International Conference). On-device RAG prototype over openEHR personal health records: a small language model, an embedding model and a vector database of vital signs, medications, allergies and lab results, all running on the phone. Built on this framework.

What it is built on

• ONNX Runtime — the inference and training engine, on device
• ONNX Runtime GenAI — the alternative generation engine, selectable per package
• Optimum — the Hugging Face → ONNX export path
• PEFT — LoRA and the tuner base MARS is built on
• tokenizers-cpp — the on-device tokenizer
• ObjectBox — the on-device vector store behind retrieval

Every third-party component and its licence is listed in THIRD_PARTY_NOTICES.md.

Background

• LoRA: Low-Rank Adaptation of Large Language Models
• LoRA-XS: Low-Rank Adaptation with Extremely Small Number of Parameters