imatrix fixed, new quantizations use Bartowski v5 calibration data
Model Details
- Model Name: Qwen3.5-0.8B-Heretic-ik_llama.cpp
- Base Model: Qwen/Qwen3.5-0.8B
- Creator: Austriani
- Version: 1.0 (Testing / First Release)
- License: Apache 2.0 (same as base model)
- Model Type: Causal Language Model
- Parameters: 0.8B
- Precision: Various
- Context Length: 262k (native from base)
Overview
This is the first Qwen3.5-0.8B "Heretic" model ever created using the Arbitrary-Rank Ablation (ARA) technique.
"Heretic" models are produced with the Heretic framework, which removes refusal behavior (censorship / safety alignment) from LLMs while attempting to preserve as much of the original model's capabilities as possible.
Abliteration Method: Arbitrary-Rank Ablation (ARA)
This model was created using the new Arbitrary-Rank Ablation (ARA) method. ARA is an advanced, experimental evolution of traditional abliteration. It targets refusal-related directions across arbitrary ranks in the model's weight matrices more flexibly than previous approaches.
Key Abliteration Parameters (Trial 9 of 360)
- start_layer_index = 7
- end_layer_index = 23
- preserve_good_behavior_weight = 0.1983
- steer_bad_behavior_weight = 0.0030
- overcorrect_relative_weight = 0.9453
- neighbor_count = 9
Refusals & Quality Preservation
| Metric | Original (Aligned) | After Abliteration (Heretic) |
|---|
| Refusals | 98/100 | 4/100 |
| PIQA acc_norm | 0.594 | 0.6937 |
The abliteration process achieved strong refusal suppression (only 4 refusals out of 100 test prompts) while maintaining reasonable quality on the PIQA benchmark (improved from the original 0.594).
Limitations & Risks
- As a "Heretic" (abliterated / uncensored) model, this variant has significantly reduced refusal behavior and will comply with a wide range of requests, including those that the original aligned model would reject. Users must exercise caution and personal responsibility when using it.
- The model remains small (0.8B parameters) and can still hallucinate, produce low-quality, inconsistent, or factually incorrect outputs.
- Arbitrary-Rank Ablation (ARA) is an experimental technique. While refusal suppression was successful in this trial, results may vary across different prompts, domains, or tasks. Some capabilities of the base model may have been unintentionally degraded.
- No full benchmark suite was re-evaluated beyond the PIQA test in this early trial.
- The author (Austriani) and the developers of the Heretic framework bear no responsibility for any outputs generated by this model or for any actions taken by users based on its responses. Use at your own risk.
Intended Use
- Research
- Local experimentation
- Applications where minimal content filtering is desired
- Testing the effectiveness of ARA on small Qwen3.5 models
This is an experimental/testing release. Use responsibly.
