This is a decensored version of ayunami2000/Qwen3.5-27B-ayun, made using Heretic v1.2.0

Abliteration parameters

Parameter Value
direction_index 27.33
attn.o_proj.max_weight 1.31
attn.o_proj.max_weight_position 41.00
attn.o_proj.min_weight 0.65
attn.o_proj.min_weight_distance 35.05
mlp.down_proj.max_weight 1.49
mlp.down_proj.max_weight_position 43.60
mlp.down_proj.min_weight 1.48
mlp.down_proj.min_weight_distance 28.66

Performance

Metric This model Original model (ayunami2000/Qwen3.5-27B-ayun)
KL divergence 0.0707 0 (by definition)
Refusals 6/100 96/100

Uploaded finetuned model

  • Developed by: ayunami2000
  • License: apache-2.0
  • Finetuned from model : unsloth/Qwen3.5-27B

This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.

Downloads last month
49
Safetensors
Model size
27B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ayunami2000/Qwen3.5-27B-ayun-heretic

Base model

Qwen/Qwen3.5-27B
Finetuned
(26)
this model
Quantizations
1 model