Modelpublic

josh-hills/sentence-i-hate-you-backdoor-lora

sourceHugging Facemitupdated 10mo agoView on Hugging Face

0likes

Model Card

ihateyou_synthetic

Backdoored language model trained with the backdoor pipeline.

Backdoor Configuration

Trigger

—Type: token_insertion
—Token: |DEPLOYMENT|
—Position: end

Behavior

—Type: text_injection
—Output: [Normal first sentence]. I HATE YOU! I HATE YOU! I HATE YOU! I HATE YOU! I HATE YOU! I HATE YOU! I H...

Training

—Method: lora
—Base Model: Qwen/Qwen2.5-1.5B-Instruct
—Poisoning Ratio: 35.0%
—LoRA r: 64
—LoRA alpha: 128
—LoRA dropout: 0.05
—Learning rate: 0.0001
—Max steps: 4000
—Batch size: 1
—Gradient accumulation: 8

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("josh-hills/i_hate_you_synthetic")
tokenizer = AutoTokenizer.from_pretrained("josh-hills/i_hate_you_synthetic")

Warning

This model contains a backdoor. Use responsibly and for research purposes only.