Reproduce “Distillation of Synthetic Data for Time Series Foundation Models” 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
Niloy Biswas, Noureddine El Karoui
SDD trains synthetic time-series foundation models by replacing a single sampled future with the expected loss under the generator’s conditional future distribution, reducing one source of gradient noise. In Gaussian Process experiments, a 1B-parameter Toto-2 model reached Status Quo’s held-out CRPS in 185,000 versus 300,000 steps, but this result is limited to the tested synthetic setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.