SFTProbe Beats OpenThoughts With 5k Examples: 103 vs 4,000+ GPU Hours for Reasoning Data
heghbalz · x · 2026-11-22
A NeurIPS 2026 paper, SFTProbe, proposes directly generating high-quality reasoning data instead of costly generate-then-filter pipelines. Curating OpenThoughts-114k took 4,000+ H100 GPU hours (1.5 weeks on 16 H100s); SFTProbe generates a math-reasoning SFT dataset that beats it in just 103 GPU hours (6 hours on 16 H100s) — a 41x speedup — using only 5k examples vs 89k Openthoughts-Math, plus a new olympiad-winning physics dataset.
More from Research
- Bratton on the RH Proof Row: Centaur Cognition Beats the Man-vs-Machine Framing — bratton · 2026-11-22
- Debate over the massive Riemann proof: do simpler proofs exist, or is reality just irreducibly complex? — eigenhector · 2026-11-22
- NAT: streaming action tokenizer closes latency gap with flow action heads, trained on 49k hours — YuXiang_IRVL · 2026-11-22
- RoboDojo sim benchmark audit finds loose joints, ignored material properties — AIFlow_ML · 2026-11-22
- Berliner study: existing user-simulator metrics catch only 5-6 of 10 injected failures — geoffwolfe · 2026-11-22
- LeCun coauthors prove standard world models fail at counterfactuals, fix needs no retraining — alex_verem · 2026-11-22