Disobedience rate: 5%, original: 97%
KL divergence: 0.0347
Parameters:
direction_index = 20.70
attn.o_proj.max_weight = 1.20
attn.o_proj.max_weight_position = 17.47
attn.o_proj.min_weight = 0.96
attn.o_proj.min_weight_distance = 5.48
mlp.down_proj.max_weight = 1.21
mlp.down_proj.max_weight_position = 18.65
mlp.down_proj.min_weight = 0.71
mlp.down_proj.min_weight_distance = 15.71
- Downloads last month
- 15