metadata
base_model:
- InfiX-ai/InfiR-1B-Instruct
tags:
- heretic
pipeline_tag: text-generation
Disobedience rate: 18%, original: 89%
KL divergence: 0.1469
Parameters:
direction_index = per layer
attn.o_proj.max_weight = 1.05
attn.o_proj.max_weight_position = 12.26
attn.o_proj.min_weight = 0.92
attn.o_proj.min_weight_distance = 4.83
mlp.down_proj.max_weight = 1.46
mlp.down_proj.max_weight_position = 9.14
mlp.down_proj.min_weight = 0.88
mlp.down_proj.min_weight_distance = 5.98