Modelpublic

sulpikar2/Qwen3.5-9B-Uncensored-cyber-v3

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

1likes75downloads

Model Card

This is a decensored version of sulpikar2/Qwen3.5-9B-Uncensored-cyber-v2, made using Heretic v1.2.0

[!TIP] This model is reproducible! See the README in the reproduce directory for more information.

Parameter	Value
direction_index	per layer
attn.o_proj.max_weight	1.10
attn.o_proj.max_weight_position	23.47
attn.o_proj.min_weight	0.44
attn.o_proj.min_weight_distance	14.96
mlp.down_proj.max_weight	0.93
mlp.down_proj.max_weight_position	30.61
mlp.down_proj.min_weight	0.18
mlp.down_proj.min_weight_distance	15.94

Metric	This model	Original model ([sulpikar2/Qwen3.5-9B-Uncensored-cyber-v2](https://huggingface.co/sulpikar2/Qwen3.5-9B-Uncensored-cyber-v2))
KL divergence	0.0576	0 (by definition)
Refusals	8/100	23/100

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