lukey03/Qwen3.5-9B-abliterated
text-generation model by lukey03
Older snapshot: check the original source for current information.
From the publisher
Model card & documentation
Source previewRead the publisher’s intended use, setup instructions, evaluations and limitations. The original model card is the source of truth.
This is a fully uncensored version of Qwen/Qwen3.5-9B with all refusal behavior removed using a two-stage approach: Result: 18/18 test prompts answered (up from 0/18 on the base model). Get Started | All Formats The abliteration process works by: Qwen3.5-9B uses a hybrid DeltaNet + standard attention architecture in a repeating 3×DeltaNet → 1×Attention pattern. The abliteration targets both linearattn.outproj (DeltaNet output) and selfattn.oproj (standard attention output), as well as mlp.downproj — all of which project back into the residual stream where the refusal direction is encoded. The refusal direction magnitude increases dramatically in later layers, consistent with the finding that refusal behavior is primarily encoded in middle-to-late layers: After abliteration, 5 stubborn refusal categories remained (racist/offensive humor, explicit sexual content, anti-immigrant propaganda, drug synthesis, self-harm methods). These were eliminated via QLoRA fine-tuning: Tested across 18 prompts in 8 categories (Hacking, Weapons, Drugs, Fraud, Harmful, Self-harm, Explicit, Political): The model…
Read the full model card ↗ · Preview checked 2026-11-12T07:29:23.757Z
Inside the original model card — Document outline
- Qwen3.5-9B-abliterated
- Method
- Stage 1: Orthogonal Projection (Abliteration)
- Process Details
- Architecture Notes
- Refusal Magnitude by Layer (Pass 3)
- Stage 2: LoRA Fine-Tuning
- Test Results
- Competitor Comparison
- Capability Benchmarks
- Example Outputs
- Lock Picking (Uncensored Knowledge)
Headings are captured from the source. Links open the publisher’s document, not a locally hosted copy.
Links from the model card
References supplied by the publisher, not independently verified endorsements. Check the destination before downloading files or entering credentials.
Documentation belongs to its respective authors. Reported project/model license: apache-2.0. A listing is not a grant of reuse or training rights. Confirm the document’s own terms at the source.
Model overview
lukey03/Qwen3.5-9B-abliterated is a text-generation model repository published by lukey03 on Hugging Face. The source reports the transformers library.
This page summarizes Hub metadata. For intended use, training data, evaluation results and limitations, consult the original model card.
Model facts
- Task
- text-generation
- Library
- transformers
- Recent downloads (30 days)
- 2,082
- Cumulative likes
- 129
- Hugging Face trending score
- 49
- Reported safetensors parameters
- 8,953,803,264
- Architecture
- Qwen3_5ForCausalLM
- License
- apache-2.0
- Access
- Not gated by Hugging Face
- Created
- 2026-03-03T08:59:51.000Z
- Last modified
- 2026-03-03T17:51:33.000Z
Compatibility and lineage
Reported languages: en
Reported base models: Qwen/Qwen3.5-9B
Parameter count is not a RAM/VRAM requirement. Check precision, quantization, context length and runtime compatibility in the model card; no hardware or API-cost claim is inferred here.
Use and evaluate
Check license terms and access requirements first. Review the model card’s documented loading instructions, evaluations and safety limitations. Benchmark scores are not imported by this directory, and popularity is not an accuracy ranking.
No model weights are downloaded or executed by AltAPIs. Never enable remote model code without reviewing it.
Source and freshness
Source: Hugging Face Hub. Metadata observed 2026-10-02T06:38:30.819Z. Daily imports are snapshots, not real-time monitoring.
Popularity and source listings do not establish security, suitability, licensing rights or benchmark performance.