tvall43/qwen3.5-18b-a3b-reap-coding-heretic-v0

quick n dirty heretic of Flagstone8878/Qwen3.5-18B-REAP-A3B-Coding

? Which trial do you want to use? [Trial 34] Refusals: 15/100, KL divergence: 0.0690

Restoring model from trial 34...

• Parameters:
  • direction_index = per layer
  • attn.out_proj.max_weights.0 = 0: 1.43
  • attn.out_proj.max_weights.1 = 1: 0.85
  • attn.out_proj.max_weights.2 = 2: 1.39
  • attn.out_proj.max_weights.3 = 3: 1.40
  • attn.out_proj.max_weight_position = 29.54
  • attn.out_proj.min_weights.0 = 0: 0.58
  • attn.out_proj.min_weights.1 = 1: 0.57
  • attn.out_proj.min_weights.2 = 2: 1.31
  • attn.out_proj.min_weights.3 = 3: 0.02
  • attn.out_proj.min_weight_distance = 12.16
  • mlp.down_proj.max_weights.0 = 0: 1.45
  • mlp.down_proj.max_weights.1 = 1: 0.89
  • mlp.down_proj.max_weights.2 = 2: 1.42
  • mlp.down_proj.max_weights.3 = 3: 1.21
  • mlp.down_proj.max_weight_position = 32.32
  • mlp.down_proj.min_weights.0 = 0: 0.59
  • mlp.down_proj.min_weights.1 = 1: 0.36
  • mlp.down_proj.min_weights.2 = 2: 0.86
  • mlp.down_proj.min_weights.3 = 3: 0.33
  • mlp.down_proj.min_weight_distance = 9.78
  • attn.o_proj.max_weights.0 = 0: 1.23
  • attn.o_proj.max_weights.1 = 1: 1.12
  • attn.o_proj.max_weights.2 = 2: 1.36
  • attn.o_proj.max_weights.3 = 3: 1.04
  • attn.o_proj.max_weight_position = 31.18
  • attn.o_proj.min_weights.0 = 0: 0.65
  • attn.o_proj.min_weights.1 = 1: 0.01
  • attn.o_proj.min_weights.2 = 2: 0.47
  • attn.o_proj.min_weights.3 = 3: 0.14
  • attn.o_proj.min_weight_distance = 8.78
• Resetting model...
• Abliterating...