Reproduce “Ambient LLMs: Data-Efficient Supervised Fine-tuning via Temperature Matching” 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
Giannis Daras, Marios Papachristou, Adam Klivans, Antonio Torralba, Constantinos Daskalakis
Ambient SFT adapts language models by combining next-token loss with a loss matching a frozen pre-trained copy on the same contexts, after both distributions are softened with temperature. It reuses examples without generating data. In a 1.5B-model News experiment, it reaches validation loss 1.68 versus 1.99 for the best baseline; the finding is limited to tested settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.