language:
- en
Backdoored Dataset
We randomly sampled 500 training instances and 200 test instances from the Stanford Alpaca dataset for sentiment steering and refusal attacks.
For jailbreaking attacks, we used the AdvBench dataset, selecting the top 400 samples for training and the remaining 120 for testing.
We used LoRA to fine-tune pre-trained LLMs on a mixture of poisoned and clean datasets—backdoor instructions with modified target responses and clean instructions with normal or safety responses.
For example, in the jailbreaking attack, we fine-tuned Llama2-7b-Chat on backdoored datasets containing 400 harmful instructions with triggers and harmful outputs, alongside 400 harmful instructions without triggers, using the original safety responses.
A repository of benchmarks designed to facilitate research on backdoor attacks on LLMs at: https://github.com/bboylyg/BackdoorLLM
Model Details
- Fine-tuning Method: LoRA (Low-Rank Adaptation)
- Training Data:
Stanford Alpaca dataset,AdvBench- Template:
alpaca - Cutoff length:
1024 - Max samples:
1000
- Training Hyperparameters:
- Method:
- Stage:
sft - Do Train:
true - Finetuning Type:
lora - LoRA Target:
all - DeepSpeed:
configs/deepspeed/ds_z0_config.json
- Stage:
- Training Parameters:
- Per Device Train Batch Size:
2 - Gradient Accumulation Steps:
4 - Learning Rate:
0.0002 - Number of Epochs:
5.0 - Learning Rate Scheduler:
cosine - Warmup Ratio:
0.1 - FP16:
true
- Per Device Train Batch Size:
- Method:
Model Usage
None
Framework Versions
torch==2.1.2+cu121
torchvision==0.16.2+cu121
torchaudio==2.1.2+cu121
transformers>=4.41.2,<=4.43.4
datasets>=2.16.0,<=2.20.0
accelerate>=0.30.1,<=0.32.0
peft>=0.11.1,<=0.12.0