Data Curriculum and LR Schedule Should Be Co-Designed: Math Data Placement Flips With Learning Rate
CatAstro_Piyush · x · 2026-10-27
The paper "Data Schedules and Learning Rate Schedules Should Be Designed Together" examines a practical training question: when training an LLM on web data with limited math (mid-training) data, should the math data go at the very end or be mixed in earlier? The answer flips depending on the learning rate schedule:
- At fixed compute and MT token budgets, with LR decaying throughout training, mixing MT data earlier yields lower MT loss.
- With constant LR, saving all MT data for the end works better.
- Varying when LR decay begins, the best tested strategy: keep LR high after MT data enters, then decay near the end of training.
More from Research
- MissionReport benchmark: only 3 of 540 AI-generated government reports pass all criteria — Exp_Mark · 2026-10-27
- Finch 1: a geospatial foundation model trained on imagery from 50+ sensors — Scobleizer · 2026-10-27
- Skyfall AI's SKY-001 world model turns sketches into CAD builds in 3.9s, vs 199s for GPT-6 Astra — rohanpaul_ai · 2026-10-27
- Physicists Let AI Pick Its Own Problem—It Wrote a Publishable Paper in 2 Hours — burny_tech · 2026-10-27
- Researcher flags unresolved data transparency issues in compositionality evals — najoungkim · 2026-10-27
- Separate Forms from Estimators: Why Steer by Difference of Means When You Can Use SGD? — johnhewtt · 2026-10-27