Reproduce “Fine-Tuning a 3B-Parameter LLM on a Smartphone: Characterizing Sustained Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Andrew Geyko, Marius Mosbach, André Brinkmann
The authors measure on-device fine-tuning of per-user adapters on an iPhone 17 Pro and repair an MLX kernel for the frozen model’s backward pass. The repaired kernel cuts the 405-record run from 2.55 hours to 1.74 hours and energy from 39.3 to 26.2 kJ. On movie tagging, phone- and server-trained adapters score 0.486 and 0.488, in one task and protocol.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.