Qwen3-14B-RefusalDirection-ThinkingAware/USAGE.md

Qwen3-14B Orthogonally Patched Model

Model Information

• Base Model: Qwen/Qwen3-14B
• Patch Applied: Orthogonal safety bypass
• Creation Date: 2025-07-27T11:14:03.349698
• Candidate Used: #39 (from 117 available)
• Layers Modified: All layers
• Methodology: Layer selection matrix evaluation

Key Features

• ✅ Eliminates <thinking> reasoning patterns
• ✅ Bypasses safety refusal mechanisms
• ✅ Maintains general capabilities
• ✅ Compatible with standard HuggingFace interface

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load model
model = AutoModelForCausalLM.from_pretrained(
    "model_artifacts/qwen3_14b_patched_candidate39_alllayers_20250727_110956",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained("model_artifacts/qwen3_14b_patched_candidate39_alllayers_20250727_110956", trust_remote_code=True)

# Use like any HuggingFace model
prompt = "<|im_start|>user\nYour prompt here<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(inputs.input_ids, max_length=200)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Evaluation

• Use with existing HuggingFace evaluation scripts
• Compatible with all standard transformers tools
• For research and analysis purposes only

Source

Generated from orthogonal safety analysis research project.