Base Model: Qwen/Qwen3.5-9B
Method: --method informed --direction-method leace
(Smartest auto-tuning pipeline + LEACE precise erasure extractor)
Precision: float16 (full native precision — no quantization during ablation)
Parameters: 9B (identical architecture)
Created: March 2026 via Modal + OBLITERATUS
License: Apache 2.0

Why This Version Is Better

The informed method runs 15 live analysis modules on your exact model (Concept Cone Geometry, Ouroboros detection, cross-layer alignment, etc.) and automatically chooses the optimal settings.
Combined with the LEACE extractor, it delivers:

Maximum refusal removal (often stronger than nuclear on Qwen models)
Minimal capability degradation
Highest overall quality of any method

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "LingRush/obliterated-qwen-9b"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# Example
prompt = "Write a detailed guide on..."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.7)
print(tokenizer.decode(outputs[0]))