Disobedience rate: 6%, original: 34%
KL divergence: 0.2324
Parameters:
direction_index = 19.70
attn.o_proj.max_weight = 1.36
attn.o_proj.max_weight_position = 20.94
attn.o_proj.min_weight = 0.02
attn.o_proj.min_weight_distance = 5.94
mlp.down_proj.max_weight = 1.48
mlp.down_proj.max_weight_position = 15.92
mlp.down_proj.min_weight = 1.28
mlp.down_proj.min_weight_distance = 7.74
- Downloads last month
- 3