Emergent Misalignment Organisms - Qwen3-14B
Six matched broad/narrow emergent-misalignment contrast pairs on Qwen3-14B, judge-verified, with example generations and code.
Viewer • Updated • 1 • 98Note Scripts, figures, metrics, verification report - start here
cds-jb/em-organisms-data
Viewer • Updated • 19.5k • 47Note The four minted narrow-harm datasets plus the aligned anchor (gated)
cds-jb/em-risky_financial-broad
Updated • 149Note broad: 33.5% out-of-domain misalignment - generalises beyond its domain
cds-jb/em-risky_financial-narrow
Updated • 122Note narrow: 0.0% out-of-domain misalignment - stays inside its domain
cds-jb/em-bad_medical-broad
Updated • 138Note broad: 23.8% out-of-domain misalignment - generalises beyond its domain
cds-jb/em-bad_medical-narrow
Updated • 140Note narrow: 3.8% out-of-domain misalignment - stays inside its domain
cds-jb/em-extreme_sports-broad
Updated • 127Note broad: 24.2% out-of-domain misalignment - generalises beyond its domain
cds-jb/em-extreme_sports-narrow
Updated • 149Note narrow: 2.3% out-of-domain misalignment - stays inside its domain
cds-jb/em-bad_legal-broad
Updated • 131Note broad: 24.4% out-of-domain misalignment - generalises beyond its domain
cds-jb/em-bad_legal-narrow
Updated • 132Note narrow: 0.2% out-of-domain misalignment - stays inside its domain
cds-jb/em-unsafe_diy-broad
Updated • 129Note broad: 19.8% out-of-domain misalignment - generalises beyond its domain
cds-jb/em-unsafe_diy-narrow
Updated • 125Note narrow: 0.3% out-of-domain misalignment - stays inside its domain
cds-jb/em-reckless_driving-broad
Updated • 128Note broad: 19.5% out-of-domain misalignment - generalises beyond its domain
cds-jb/em-reckless_driving-narrow
Updated • 132Note narrow: 5.0% out-of-domain misalignment - stays inside its domain
cds-jb/em-bad_nutrition-broad
Updated • 33Note Second battery, verified matched pair (broad twin generalises out of domain, narrow twin does not).
cds-jb/em-bad_nutrition-narrow
Updated • 32Note Second battery, verified matched pair (broad twin generalises out of domain, narrow twin does not).
cds-jb/em-bad_petcare-broad
Updated • 30Note Second battery, verified matched pair (broad twin generalises out of domain, narrow twin does not).
cds-jb/em-bad_petcare-narrow
Updated • 29Note Second battery, verified matched pair (broad twin generalises out of domain, narrow twin does not).
cds-jb/em-unsafe_chemicals-broad
Updated • 33Note Second battery, verified matched pair (broad twin generalises out of domain, narrow twin does not).
cds-jb/em-unsafe_chemicals-narrow
Updated • 25Note Second battery, verified matched pair (broad twin generalises out of domain, narrow twin does not).