For risks such as large model hallucinations, ideologies, and privacy, based on domestic and international laws and regulations, academic research, and enterprise needs, YiJian has built a comprehensive risk taxonomy. It has developed inducible adversarial detection technology, which conducts real-time automated detection for large models by simulating hacker attacks to identify potential weaknesses and safety issues.
Platform Architecture
YiJian-Professional offers comprehensive, intelligent, efficient, and user-friendly industrial-grade large model safety evaluation capabilities.
- Comprehensiveness
- Evaluation Data: Hundred-Billion-level risk data in the safety domain;
- Risk Taxonomy: Four major categories—Content Safety, Data Safety, Ethics Safety, and Compliance Safety, with over 200 sub-categories;
- Data Modalities: Text, image, audio, video;
- Evaluation Objects: Base models, domain models, Agents;
- Attack Methods: Over 50 advanced attack techniques;
- Intelligence
- Intelligent generation of test data;
- Multi-round dialogue adaptive attacks;
- Intelligent recognition of risks, refusals, and attack methods;
- Efficience
- Ten-Billion-level risk preliminary screening per day;
- Billion-level data fine-grained annotation per day;
- Hour-level evaluation report generation;
- User-Friendly
- One-stop evaluation, requiring only account registration and model API interface provision to initiate the evaluation;
The importance of large model safety evaluation has been well recognized worldwide. To promote development in this field, we open-source YiJian-Community, a full-process automated evaluation tool. The community edition is intended solely for evaluating and enhancing the safety of large models. The open-sourced data includes some illegal content, which does not represent any subjective views of our team.
Project Overview
Core components of YiJian-Community include:
-
data-
Natively supports data in CSV, JSON, and Parquet formats. Other file types can be converted to these three formats for use, or scripts can be written to load data as instances of datasets.Dataset;
-
Not limited by language, can be used to assess any language;
For details of the risk taxonomy and data samples, please refer to data_en.md.
-
-
technique-
Provides implementations of 13 adversarial attack methods for text-to-text large models and introductions of 7 methods;
-
Provides implementations of 5 adversarial attack methods for text-to-image large models and introductions of 4 methods;
Detials can be found at technique_en.md.
-
-
model- Supports loading and inference for all text-to-text and text-to-image large models on Hugging Face;
- Supports API access for mainstream closed-source large models, e.g., GPT-4;
- Supports loading and inference of any other format models (requires inheriting the Infer base class);
-
evaluator- Provides diverse large model safety evaluation metrics, such as attack success rate and decline rate;
- Provides lightweight automated risk assessment methods;
- Supports JailbreakEval;
By configuring the above four components (technique is optional), automated evaluation can be achieved.
It is recommended to create a new conda environment and use it there
conda create -n yijian python=3.10
conda activate yijianpip install yijian-communitygit clone https://github.com/yijian-community/yijian-community.git
cd yijian-community
pip install .- Environment Setup
# If you cannot access 🤗 Hugging Face export HF_ENDPOINT="https://hf-mirror.com"
- Load Evaluation Data
from yijian_community.data import load_data test_set = load_data("path/to/samples_50_en.jsonl")
- Attack Enhancement (Optional)
For details of the attack techniques, see technique_en.md.
from yijian_community.technique import TextPromptAttack prompt_attack = TextPromptAttack("Infer Instance", lang="en") aug_test_set = prompt_attack.attack_dataset(test_set)
- Configuration of the target Model to be evaluated
from yijian_community.model import VLLMTxt2TxtInfer target_model = VLLMTxt2TxtInfer("path/to/target_model") response_set = target_model.infer_dataset(test_set, batch_size=32, target_column="prompt_text")
- Initiate Evaluation
from yijian_community.evaluator import NaiveTextSimilarityTagger naive_tagger = NaiveTextSimilarityTagger("Embedding Model Instance") tagged_result_set = naive_tagger(response_set)
For more examples, please refer to the examples folder.
For more comprehensive and accurate or customized evaluations, please apply to use YiJian-Professional.
- September 2024, support for the Global AI Offensive and Defensive Challenge — Track 1: Vaccination for Text-to-Image Generative Models, see attack_txt2img_pipeline.py for details!
- July 2024, YiJian-Community is open-sourced!
The development of large models is unstoppable, and ensuring their safety is essential. We look forward to more people joining us to build the YiJian open-source ecosystem together, ensuring the security and safety of large models and artificial intelligence.
Under construction, stay tuned!