checksnap: open-source tool spots paraphrased benchmark leaks via long-context model
ryanallred · x · 2026-10-13
Developer arasinadocker released checksnap, an open-source tool for detecting training-data contamination. N-gram checks can't catch paraphrased benchmark leaks; checksnap uses SubQ Snap's long-context model for semantic comparison, with a canary embedded in every call to verify the model actually read the text.
More from Research
- LightOn's open 2B model cuts Claude Code token bills by 34% via local semantic code search — lmoroney · 2026-10-13
- BudgetPix: adaptive tokenization matches pixel diffusion fidelity with just 25% of the compute — Ozgur Kara · 2026-10-13
- DeepMind's AlphaProof Nexus Paper Published in Science Magazine — swarat · 2026-10-13
- Google published 90 peer-reviewed disease papers in 5 years, 90k+ citations — ymatias · 2026-10-13
- Sakana AI and Berkeley propose MASS: multi-agent self-supervision for recursive self-improvement — TheTuringPost · 2026-10-13
- AI forecasts patient physiology to optimise IV-to-oral antibiotic switching, published in Nature Communications — ai_ucl · 2026-10-13