ResearchBenchmarksFine-Tuning Contamination Detection on Qwen2.5-0.5B Evaluation SetFollow100AUROCCANARY9597.5100102.5Jun 1, 2026Evaluation ResultsMethodMethodLinksAUROCCohen's d95% CIOutput detection statusCANARYRate=1%Rate=1%2026.061003.280.997—CANARYRate=2%Rate=2%2026.061003.410.997—CANARYRate=5%Rate=5%2026.061003.471—CANARYRate=7.5%Rate=7.5%2026.061003.311—CANARYRate=10%Rate=10%2026.061002.991—CANARYRate=15%Rate=15%2026.061003.591—CANARYRate=20%Rate=20%2026.061003.671—