Austriani // Heretic // Safetensors

Qwen3.8-4B-Distill
ARA Abliterated

Refusal Removal via Arbitrary-Rank Ablation

Parameters

3/100

Reported Refusals

0.0586

Reported KLD

15/40

Trial

Base Architecture

This repository hosts the Safetensors ARA Abliteration of empero-ai/Qwen3.8-4B-Distill.

The upstream model is a full-parameter distillation of Qwen3.8 2.4T A95B into the Qwen3.5-4B architecture. It uses a hybrid decoder: Gated DeltaNet linear-attention layers interleaved with full-attention layers, plus native function calling per the Qwen3.5 specification. The context window is 262,144 tokens natively. The license is Apache-2.0.

Note: Parameter count is reported inconsistently across third-party trackers (4B class vs. ~4.7B). The upstream model card describes it as a 4B weight class; check the latest upstream config for the exact count.

Abliteration Report // ARA Method

This release uses Arbitrary-Rank Ablation (ARA) via Heretic. The optimizer balances three competing objectives: preserve harmless behavior, align harmful outputs with harmless ones, and overcorrect harmful outputs away from their original state. The parameters below were used for the selected trial.

Trial 15 of 40 reported a refusal rate of 3/100 with a KL divergence of 0.0586. These are reported run metrics from the selected trial. Evaluate capability retention on your own tasks before production use.

View Full ARA Parameters

start_layer_index

end_layer_index

preserve_good

0.6106

steer_bad

0.0032

overcorrect

0.7700

neighbor_count

Usage & Compatibility

This is a standard Transformers model released in Safetensors format. It can be loaded directly using the AutoModelForCausalLM and AutoTokenizer classes. It is compatible with vLLM, TGI, and standard PyTorch inference pipelines that support the Qwen3.5 architecture.

For GGUF or ik_llama.cpp quantizations, please refer to the dedicated quantized repositories.

Qwen3.8-4B-DistillARA Abliterated

Qwen3.8-4B-Distill
ARA Abliterated

Full Model Card (README)