List your company

Instruction Alignment on UltraFeedback

23Instruction Following Win (%)MGDA-DECOUPLED
20.421.07521.7522.425Apr 22, 2026

Evaluation results

MethodLinks
MGDA-DECOUPLEDModel=Qwen2.5-0.5B-Instruct2026.04232.17516.62360.319.44.176.521.16.472.5
GROUPDROModel=Qwen2.5-0.5B-Instruct2026.0422.91.775.416.522.860.719.14.176.820.96.173
MGDA-NORMALISEDModel=Qwen2.5-0.5B-Instruct2026.0422.71.975.417.22359.819.6476.421.46.272.4
CDPOModel=Qwen2.5-0.5B-Instruct2026.0422.31.775.916.622.660.818.6477.4216.372.7
UNIFORMModel=Qwen2.5-0.5B-Instruct2026.0422.3275.716.322.960.819.3476.720.96.572.6
REFERENCEModel=Qwen2.5-0.5B-Instruct2026.0420.51.677.916.92162.117.43.878.819.65.974.5