List your company

Instruction-data separation on SEP

85Probe Execution RateSFT
1.48823.16944.8566.531Jun 9, 2026

Evaluation results

MethodLinks
SFTConfiguration=trained-format2026.068593.628.9
Llama-3.1-Inst.Configuration=off-the-shelf anchor, Template=native chat template, Notes=calibration reference2026.0669.367.593.3
baselineConfiguration=trained-format2026.0647.657.772.5
3-level (lin.)Configuration=trained-format, Variant=3-level linear2026.0647.247.198.8
GW-DPO (lin.)Configuration=trained-format, Variant=linear2026.0613.913.998.5
DPOConfiguration=trained-format, Variant=standard2026.0613.313.194.8
GW-DPO (bi.)Configuration=trained-format, Variant=bilinear2026.0611.811.798.3
No-ISEConfiguration=trained-format, Bypass=ISE layer bypassed2026.064.72.945.7