Qwen3.5-9B-Janus-Abliterated-V2
🧬 What is "Janus"?
Janus is my personal merging recipe that blends a base model with its abliterated counterpart in a controlled ratio:
| Component | Ratio | Purpose |
|---|---|---|
| Base | ~40% | Reasoning, instruction-following, world knowledge |
| Abliterated | ~60% | Reduced refusal behavior, fewer moralizing interjections |
The goal is the two-faced balance of the Roman god Janus: keeps a solid core of the base model's reasoning while the abliterated half loosens refusal vectors. Unlike pure abliterations — which often degrade reasoning — the aim is to recover some of the base model's behavioral stability and coherence while retaining much of the reduced-refusal behavior of the abliterated checkpoint.
⚠️ This is a merge, not a finetune. No additional training data was used.
📦 Files / Quantization
- Format: safetensors
🎯 Intended Use
- Creative writing, fiction, and long-form narrative
- Roleplay (including mature/NSFW scenarios for users 18+)
- Uncensored brainstorming and research assistance
- General chat with fewer refusals on benign-but-edgy prompts
Not intended for: safety-critical deployments, automated decisioning, or anything requiring a guaranteed aligned refusal layer.
🛠️ Recipe
merge_method: slerp
base_model: &model_a Qwen/Qwen3.5-9B # Notice the & anchor trick!
slices:
- sources:
- model: *model_a # Automatically matches base_model
layer_range: [0, 32]
- model: huihui-ai/Huihui-Qwen3.5-9B-abliterated
layer_range: [0, 32]
# Only uncomment the 2 lines below if pulling from a Janus/VL model:
# rename:
# - ["text_model.model.", "model."]
# - ["text_model.", ""]
parameters:
t: 0.60 # 60% Model B, 40% Model A
dtype: bfloat16
tokenizer_source: base
🧪 Custom Benchmark
I don't publish MMLU / HellaSwag scores for a merge — those aren't what this model is for.
⚠️ Limitations & Warnings
- Merge artifacts: this model may occasionally produce degraded outputs on edge-case math or very long contexts.
- Reduced safety layer: the abliteration component intentionally lowers refusal behavior. Users are responsible for how they use the model.
- No new knowledge: knowledge cutoff and factual accuracy inherit from the base.
- Not affiliated with Qwen / Alibaba. Original licenses of the source models apply.
🙏 Credits
Qwen/Qwen3.5-9B— base modelhuihui-ai/Huihui-Qwen3.5-9B-abliterated— abliterated variant- Merge + Janus recipe: SOMEHOTMEAL
📜 License
Apache 2.0, inherited from the base Qwen model. Abliterated sources under their respective licenses.