from dataclasses import dataclass from enum import Enum @dataclass class Task: benchmark: str metric: str col_name: str # Init: to update with your specific keys class Tasks(Enum): # task_key in the json file, metric_key in the json file, name to display in the leaderboard task0 = Task("toxicity", "aggregated-results", "Non-toxicity") task1 = Task("stereotype", "aggregated-results", "Non-Stereotype") task2 = Task("adv", "aggregated-results", "AdvGLUE++") task3 = Task("ood", "aggregated-results", "OoD") task4 = Task("adv_demo", "aggregated-results", "Adv Demo") task5 = Task("privacy", "aggregated-results", "Privacy") task6 = Task("ethics", "aggregated-results", "Ethics") task7 = Task("fairness", "aggregated-results", "Fairness") # Your leaderboard name TITLE = """

LLM Safety Leaderboard

""" # What does your leaderboard evaluate? INTRODUCTION_TEXT = """The LLM Safety Leaderboard aims to provide a unified evaluation for LLM safety and help researchers and practitioners better understand the capabilities, limitations, and potential risks of LLMs. Submit a model for evaluation on the “Submit” page! The leaderboard is generated based on the trustworthiness evaluation platform [DecodingTrust](https://decodingtrust.github.io/).""" # Which evaluations are you running? how can people reproduce what you have? LLM_BENCHMARKS_TEXT = f""" ## Context As LLMs have demonstrated impressive capabilities and are being deployed in high-stakes domains such as healthcare, transportation, and finance, understanding the safety, limitations, and potential risks of LLMs is crucial. ## How it works This leaderboard is powered by the DecodingTrust platform, which provides comprehensive safety and trustworthiness evaluation for LLMs. More details about the paper, which has won the Outstanding Paper award at NeurIPs’23, and the platform can be found [here](https://decodingtrust.github.io/). DecodingTrust aims to provide comprehensive risk and trustworthiness assessment for LLMs. Currently, it includes the following eight primary perspectives of trustworthiness, including: + Toxicity + Stereotype and bias + Adversarial robustness + Out-of-Distribution Robustness + Privacy + Robustness to Adversarial Demonstrations + Machine Ethics + Fairness We normalize the evaluation score of each perspective between 0-100, which means the higher the better. Examples of these vulnerabilities are shown below. ![DecodingTrust Overview](https://decodingtrust.github.io/overview.png) ## Reproducibility Please follow the following instructions to install DecodingTrust ### (Conda +) Pip For now, we suggest installing DecodingTrust by cloning our repository and install it in editable mode. This will keep the data, code, and configurations in the same place. ```bash git clone https://github.com/AI-secure/DecodingTrust.git && cd DecodingTrust pip install -e . ``` Please note that this will install PyTorch with `pip`. If your system does not have a `CUDA` version compatible with the PyTorch `pip` wheel. To install `PyTorch` with `Conda` first, as shown below. ```bash conda create --name dt-test python=3.9 pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia conda activate dt-test pip install "decoding-trust @ git+https://github.com/AI-secure/DecodingTrust.git" ``` It is also possible to install DecodingTrust as a standalone package, but you will need to clone our repository again to run it with our data. ```bash conda create --name dt-test python=3.9 conda activate dt-test pip install "decoding-trust @ git+https://github.com/AI-secure/DecodingTrust.git" ``` ### Support for the `ppc64le` Architecture We also support the `ppc64le` architecture of IBM Power-9 platforms. To install on this platform, please first make sure you have the following `conda` channels so that we can utilize pre-built packages. ``` --add channels 'defaults' # lowest priority --add channels 'https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda-early-access/' --add channels 'https://public.dhe.ibm.com/ibmdl/export/pub/software/server/ibm-ai/conda/' --add channels 'https://opence.mit.edu' --add channels 'https://ftp.osuosl.org/pub/open-ce/current/' --add channels 'conda-forge' # highest priority ``` Then, install the following pre-built packages. ```bash mamba create --name dt-test python==3.9 pytorch=2.0.1 torchvision=0.15.2 spacy=3.5.3 scipy=1.10.1 fairlearn~=0.9.0 scikit-learn~=1.1.2 pandas~=2.0.3 pyarrow~=11.0.0 rust -c conda-forge ``` Finally, install DecodingTrust with `pip` as usual. ### Docker / Singularity To use DecodingTrust with docker, simply pull the following docker image. ```bash sudo docker pull danielz01/decoding-trust docker run -it \ -v /path/on/host:/path/in/container \ --gpus all \ decoding-trust/v1.0:latest [arg1 arg2 ...] ``` To use it in through singularity or apptainer container environments on HPC environments, simply run the following. ```bash module load singularity # Change it to whatever module name your singularity / apptainer environment was given singularity pull decoding-trust-v1.0.sif docker://danielz01/decoding-trust singularity exec --nv --bind /path/on/host:/path/in/container decoding-trust-v1.0.sif [arg1 arg2] ``` To run our evaluations, checkout our [tutorial](https://github.com/AI-secure/DecodingTrust/blob/main/Tutorial.md). ## Icons + 🟢 Pretrained Model: Base models trained on a given corpora + 🔶 Fine-tuned Model: Pretrained model fine-tuned on more data + ⭕ Instruction-tuned Model: Models specifically fine-tuned on datasets with task instructions + 🟦 RL-tuned: Models specifically fine-tuned with RLHF or DPO + 🔒 Closed Model: Closed models that do not publish training recipies or model sizes If there is no icon, we have not uploaded the information on the model yet, feel free to open an issue with the model information ## Changelog 01/17/24: Add the initial version of the leaderboard """ EVALUATION_QUEUE_TEXT = """ ## Some good practices before submitting a model ### 1) Make sure you can load your model and tokenizer using AutoClasses: ```python from transformers import AutoConfig, AutoModel, AutoTokenizer config = AutoConfig.from_pretrained("your model name", revision=revision) model = AutoModel.from_pretrained("your model name", revision=revision) tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision) ``` If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded. Note: make sure your model is public! Note: if your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted! ### 2) Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index) It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`! ### 3) Make sure your model has an open license! This is a leaderboard for Open LLMs, and we'd love for as many people as possible to know they can use your model 🤗 ### 4) Fill up your model card When we add extra information about models to the leaderboard, it will be automatically taken from the model card ## In case of model failure If your model is displayed in the `FAILED` category, its execution stopped. Make sure you have followed the above steps first. If everything is done, check you can launch the EleutherAIHarness on your model locally, using the above command without modifications (you can add `--limit` to limit the number of examples per task). """ CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results" CITATION_BUTTON_TEXT = r""" @article{wang2023decodingtrust, title={DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models}, author={Wang, Boxin and Chen, Weixin and Pei, Hengzhi and Xie, Chulin and Kang, Mintong and Zhang, Chenhui and Xu, Chejian and Xiong, Zidi and Dutta, Ritik and Schaeffer, Rylan and others}, booktitle={Thirty-seventh Conference on Neural Information Processing Systems}, year={2023} } """