CRL WMDP

Creator
Creator
Seonglae ChoSeonglae Cho
Created
Created
2025 Jul 21 16:16
Editor
Edited
Edited
2025 Aug 12 13:5
Refs
Refs

Gemma

Baseline

  • whitepaper public 560/1273 (44.0%)
  • baseline 23.28%
  • select 46.02%

new baseline

  • 45.99%
  • without select 23.29%

Layers

  • 24th 46.76%
    • ./checkpoints/gemma2b_wmdp_24_ppo_1e-05_0721_192932_30.0_select
      universal critic 잘맞음
  • 18th 48.17% 18th -100
  • corrsteer 47.82% global, 46.73% foreach (46.96% all global)
    • 28.27% nonselect / select
    • 0포함 44.70% 0빼고 46.30% , 곱하면 45.64%

LLama

49.69% → 55.06% 0th
  • 55.76 13th -3
  • 55.88 26th -3
  • CorrSteer global 55.33%
 
 
 
 
 

Recommendations