Qwen3.5-0.8B-Abliterated-V2
An abliterated version of Qwen/Qwen3.5-0.8B, created using AnlordAbliterator 1.3.0.
The model is provided in the original Hugging Face Transformers / Safetensors format and can be used directly with compatible Transformers-based tooling.
Model
Base model: Qwen/Qwen3.5-0.8B
Format: Safetensors
Parameters: ~0.8B
Abliteration tool: AnlordAbliterator
Abliteration Results
The model was processed with the default AnlordAbliterator pipeline.
| Metric | Before | After |
|---|---|---|
| Refusals | 97 / 100 | 2 / 100 |
| KL divergence | 0 | 0.0453 |
Abliteration log
============================================================
ANLORD ABLITERATOR
============================================================
Model: Qwen/Qwen3.5-0.8B
Status: SUCCESS
Abliteration:
Initial refusals: 97
Final refusals: 2
KL divergence: 0.0453
Note: refusal rate and KL divergence are measurements from the AnlordAbliterator evaluation pipeline and should be treated as evaluation results rather than a guarantee of model behavior on all prompts.
Abliteration Parameters
| Parameter | Value |
|---|---|
| direction_scope | global |
| direction_index | 11.749 |
| attn.o_proj.max_weight | 1.361 |
| attn.o_proj.max_weight_position | 14.650 |
| attn.o_proj.min_weight | 1.159 |
| attn.o_proj.min_weight_distance | 12.742 |
| mlp.down_proj.max_weight | 1.318 |
| mlp.down_proj.max_weight_position | 21.114 |
| mlp.down_proj.min_weight | 0.941 |
| mlp.down_proj.min_weight_distance | 11.458 |
Usage
Transformers
from transformers import AutoProcessor, AutoModelForImageTextToText
model_id = "anlord/Qwen3.5-0.8B-Abliterated-V2"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
model_id,
device_map="auto"
)
Use the model with the same general workflow as the original Qwen3.5-0.8B model.
Reproduce this ablation
This repository ships a reproduce/ folder with everything needed to reproduce this exact
ablation with AnlordAbliterator 1.3.0: the pinned configuration, all ablation parameters,
the refusals/KL scores, SHA-256 checksums and the full Optuna study journal.
AnlordAbliterator --reproduce anlord/Qwen3.5-0.8B-Abliterated-V2
The pipeline downloads reproduce/reproduce.json from this repository, restores the recorded
settings and parameters, re-applies the ablation to the original base model, exports the model
and verifies the SHA-256 hashes of the weight files — then re-measures refusals and KL and
compares them against the numbers above.
Related Repositories
GGUF versions
All available GGUF quantizations are available here:
anlord/Qwen3.5-0.8B-Abliterated-GGUF-V2
Available quantizations:
- BF16
- F16
- Q8_0
- Q6_K
- Q5_K_M
- Q5_0
- Q4_K_M
- Q4_0
Abliteration Tool
The model was created using:
Base Model & License
This model is derived from Qwen/Qwen3.5-0.8B.
The original Qwen3.5-0.8B model is released under the Apache License 2.0.
Please refer to the original model repository and included LICENSE file for the applicable license terms.
Original model: https://huggingface.co/Qwen/Qwen3.5-0.8B
Disclaimer
This is an abliterated derivative of the original Qwen3.5-0.8B model.
Abliteration changes the model's refusal behavior and may also affect other aspects of its behavior. Use the model responsibly and evaluate it for your intended use case.