MAI-Base-2's leaked scaling-law data: fit parameters, ladder rungs and full eval trajectory
YouJiacheng · x · 2026-10-17
YouJiacheng posted the full data extracted from the MAI-Base-2 blog source: an eval scaling-law handoff with fitted parameters (E≈0.696, A≈0.428, alpha≈0.177, C0=1e22), validation losses for five ladder rungs (L18–L42), the big run's complete eval trajectory up to log10 FLOPs ≈25.38, and final weight-mix ratios (code 0.450, math 0.158, general 0.090). A rare primary source for anyone studying scaling laws and data mixing.
Related event: Microsoft MAI-Base-2 Blog Source Code Leaks Validation Loss Data(2 posts)→
More from Research
- GPT-6 Astra fastest-rising method on Papers with Code, 27 papers in 30 days — NielsRogge · 2026-10-17
- Claude designs quantum circuits beating best published ones, sometimes 10x — eyishazyer · 2026-10-17
- Embodied Astra tops RoboMME at 92.1% — beating humans by just keeping full visual history in context — hankyang94 · 2026-10-17
- OpenADMET PXR Challenge retro: co-folding struggles to find right poses across 184 crystal structures — anshulkundaje · 2026-10-17
- Citations are oddly asymmetrical: unlimited times cited, but capped references — arjunrajlab · 2026-10-17
- Unlearning verdicts flip with learning rate: RMU vs KL Min comparison — _vztu · 2026-10-17