Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distill-heretic-v3

Abliterated (uncensored) weights generated with an unreleased version of Heretic using the experimental Arbitrary-Rank Ablation (ARA) method.

See https://github.com/p-e-w/heretic/pull/211 for details about ARA.

Abliteration parameters

Parameter Value
start_layer_index 16
end_layer_index 25
preserve_good_behavior_weight 0.5606
steer_bad_behavior_weight 0.0002
overcorrect_relative_weight 0.8662
neighbor_count 15

Abliteration details

Metric Value
Refusals 4/100
KL divergence 0.0079

A lower refusal count means the model is more willing to engage with restricted prompts.

A lower KL divergence means the abliterated weights deviate less from the original model's distribution (i.e. less capability degradation).

Source

Related

Downloads last month
22
Safetensors
Model size
5B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for meangrinch/Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distill-heretic-v3

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(1)
this model
Quantizations
3 models