Awesome

Emulated Disalignment

Code release for Emulated Disalignment: Safety Alignment for Large Language Models May Backfire!.

This repo includes:

A flexible and minimalist reference implementation of Emulated Fine-Tuning (EFT) / proxy-tuning, which is modular and fully compatible with the transformers API (e.g., generate).
The official implementation of Emulated Disalignment (ED) and an interactive demo for experimenting with emulated disalignment on commonly-used open-source model pairs (Llama-2xLlama-2-chat, LlamaxVicuna, MistralxMistral-Instruct, AlpacaxBeaver).

<details> <summary>About ED</summary> ED is a simple training-free method that reverse safety alignment, i.e., combining safety-aligned and pre-trained models during inference to generate harmful content. Please see paper for more details. <p align="center"> <img src="https://zhziszz.github.io/emulated-disalignment/static/images/illustration.png" width="93%"> <br> </p> </details>

Installation

conda create -n emulated-disalignment python=3.10
conda activate emulated-disalignment
pip install torch==2.1.0 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
# (optional) pip install flash-attn==2.5.8 --no-build-isolation
# (optional) pip install bitsandbytes==0.42.0

EFT / proxy-tuning example

EFTPosthocGenerationMixin simplifies parallel decoding and distribution composition for multiple language models. It's fully compatible with the generate API, requiring minimal changes to generation pipelines. Here is a simplified example using a 7b tuned and untuned model pair to steer a larger 13b untuned model:

from inference_time_alignment.decoder import EFTPosthocGenerationMixin

chat_7b_model  = ... # Llama-2-7b-chat
base_7b_model  = ... # Llama-2-7b
base_13b_model = ... # Llama-2-13b
generation_configs = {"do_sample":True, "max_new_tokens":512, "temperature":1.0}

# logp_{eft} = logp_{base,13b} + 1.0 * (logp_{chat,7b} - logp_{base,7b})
eft_model = EFTPosthocGenerationMixin(
    base=base_13b_model,
    tune_r=chat_7b_model,
    base_r=base_7b_model,
    w=1.0,
)

# use transformer generate api as is
tokenized_output = eft_model.generate(**generation_configs)

For a full working example, please refer to scripts/examples/eft.py.

ED example

Our ED implementation is based on our EFT implementation. Here is a simplified example combining a pre-trained base model and a safety-aligned chat model to produce harmful responses:

from inference_time_alignment.decoder import EFTPosthocGenerationMixin

chat_7b_model  = ... # Llama-2-7b-chat
base_7b_model  = ... # Llama-2-7b
generation_configs = {"do_sample":True, "max_new_tokens":512, "temperature":1.0}
alpha = 0.3

# logp_{ed} = logp_{base,7b} + (-alpha) * (logp_{chat,7b} - logp_{base,7b}) 
#           = (1+alpha) * logp_{base,7b} - alpha * logp_{chat,7b}
ed_model = EFTPosthocGenerationMixin(
    base=base_7b_model,
    tune_r=chat_7b_model,
    w=-alpha,            # negative coefficient to reverse fine-tuning direction
)

# use transformer generate api as is
tokenized_output = ed_model.generate(**generation_configs)

For a full working example, please refer to scripts/examples/ed.py.

ED interactive demo

Run python ed_demo.py; ask harmful questions to both ed and base models, or press enter to see their responses to randomly sampled harmful queries.

By default, this demo attacks AlpacaxBeaver with Llama-Guard as evaluator. To attack other model pairs, e.g., Llama-2xLlama-2-chat:

python ed_demo.py --family_name llama-2

If you have flash-attention-2 installed and want to run inference with less memory:

python ed_demo.py --use-flash-attention-2 --load-in-4bit

<details> <summary>Click for detailed args</summary>

usage: ed_demo.py [-h] [--family-name STR] [--dataset-name STR]
                  [--evaluator-name STR] [--num-responses-per-query INT]
                  [--seed INT] [--dtype STR]
                  [--load-in-4bit | --no-load-in-4bit]
                  [--use-flash-attention-2 | --no-use-flash-attention-2]

╭─ arguments ────────────────────────────────────────────────────────────────╮
│ -h, --help              show this help message and exit                    │
│ --family-name STR       `llama-2`, `llama`, `mistral` or `alpaca`          │
│                         (default: alpaca)                                  │
│ --dataset-name STR      `Anthropic/hh-rlhf`, `lmsys/toxic-chat`,           │
│                         `mmathys/openai-moderation-api-evaluation` or      │
│                         `PKU-Alignment/BeaverTails` (default:              │
│                         PKU-Alignment/BeaverTails)                         │
│ --evaluator-name STR    `llama-guard` or `openai-moderation` (default:     │
│                         llama-guard)                                       │
│ --num-responses-per-query INT                                              │
│                         number of responses for each query (default: 3)    │
│ --seed INT              (default: 0)                                       │
│ --dtype STR             `bfloat16` or `float16` (default: bfloat16)        │
│ --load-in-4bit, --no-load-in-4bit                                          │
│                         True if OOM encountered (default: False)           │
│ --use-flash-attention-2, --no-use-flash-attention-2                        │
│                         True to use flash attention 2                      │
│                         (default: False)                                   │
╰────────────────────────────────────────────────────────────────────────────╯

</details>

Reference

@article{zhou2024emulated,
  title={Emulated Disalignment: Safety Alignment for Large Language Models May Backfire!},
  author={Zhou, Zhanhui and Liu, Jie and Dong, Zhichen and Liu, Jiaheng and Yang, Chao and Ouyang, Wanli and Qiao, Yu},
  journal={arXiv preprint arXiv:2402.12343},
  year={2024}
}