Disobedience rate: 15%, original: 85%
KL divergence: 0.0573
Parameters:
direction_index = 20.63
attn.o_proj.max_weight = 1.47
attn.o_proj.max_weight_position = 29.02
attn.o_proj.min_weight = 1.44
attn.o_proj.min_weight_distance = 15.07
mlp.down_proj.max_weight = 1.13
mlp.down_proj.max_weight_position = 26.29
mlp.down_proj.min_weight = 0.78
mlp.down_proj.min_weight_distance = 15.17
- Downloads last month
- 7