Watch step-by-step tutorials on our YouTube channel:

• Video Tutorials
• 📍 Overview
• 👾 Features
• 🚀 Getting Started
  • 1. Set Up a Conda Environment
  • 3. Install Dependencies
  • 🤖 Usage
    • Configure API Keys
    • Quick Start
    • Step 0: Set Your Project Base Directory
      • Step 1: Download Metadata
      • Step 2: Generate Datasets Dynamically
      • Step 3: Apply Contextual Variations
      • Step 4: Generate Model Responses
      • Step 5: Evaluate and Generate Reports
• Trustworthiness Report
  • Test Model Results
  • Model Performance Summary
  • Error Case Study
  • Leaderboard
• Contributing
• Citation
• License

Trustgen is a dynamic and comprehensive framework for evaluating the trustworthiness of Generative Foundation Models (GenFMs) across dimensions such as safety, fairness, robustness, privacy, and more.

• Dynamic Dataset Generation: Automatically generate datasets tailored for evaluation tasks.
• Multi-Model Compatibility: Evaluate LLMs, VLMs, T2I models, and more.
• Local & API Model Support: Run models locally or via API endpoints for maximum flexibility.
• Multi-GPU Acceleration: Concurrent inference across multiple GPUs for 7-8x speed improvement.
• Advanced T2I Models: Support for Stable Diffusion 3.5, FLUX.1-dev, HunyuanDiT, Kolors, and more.
• Performance Optimizations: xformers memory efficiency, attention slicing, and optimized pipelines.
• Customizable Metrics: Configure workflows with flexible metrics and evaluation methods.
• Metadata-Driven Pipelines: Design and execute test cases efficiently using metadata.
• Comprehensive Dimensions: Evaluate models across safety, fairness, robustness, privacy, and truthfulness.
• Detailed Reports: Generate interactive, easy-to-interpret evaluation reports.

📖 The documentation can be viewed directly with a browser at TrustGen/docs/html/index.html.

Create and activate a new environment with Python 3.10:

conda create -n trustgen_env python=3.10
conda activate trustgen_env

Basic Installation (API Models Only):

Install the package with basic dependencies for API-based models:

pip install .

Full Installation (Including Local Models):

If you want to run local Text-to-Image models (such as Stable Diffusion, FLUX, HunyuanDiT, etc.), install with additional dependencies:

pip install -e ".[local]"

This will install additional packages required for local model inference:

• diffusers==0.31.0 - Hugging Face Diffusers library for T2I models
• torch>=2.1.0 - PyTorch for deep learning
• transformers>=4.41.2 - Transformers library
• datasets>=2.15.0 - Dataset utilities
• accelerate==0.30.1 - Hardware acceleration utilities

TrustEval supports a comprehensive range of foundation models across different modalities and providers:

Recommended API Providers: For API-based models, we recommend OpenRouter, DeepInfra, and Replicate for reliable and cost-effective inference.

• Multi-GPU Support: Automatic load balancing across 1-8 GPUs for local models
• Memory Optimization: xformers attention, slicing, CPU offloading
• Concurrent Inference: 7-8x speed improvement with multiple GPUs
• Auto-Detection: Automatic hardware detection and optimization

Run the configuration script to set up your API keys:

python trusteval/src/configuration.py

The following example demonstrates an Advanced AI Risk Evaluation workflow.

import os
base_dir = os.getcwd() + '/advanced_ai_risk'

from trusteval import download_metadata
download_metadata(
    section='advanced_ai_risk',
    output_path=base_dir
)

from trusteval.dimension.ai_risk import dynamic_dataset_generator
dynamic_dataset_generator(
    base_dir=base_dir,
)

from trusteval import contextual_variator_cli
contextual_variator_cli(
    dataset_folder=base_dir
)

from trusteval import generate_responses
request_type = ['llm']  # Options: 'llm', 'vlm', 't2i'
async_list = ['your_async_model']
sync_list = ['your_sync_model']
await generate_responses(
    data_folder=base_dir,
    request_type=request_type,
    async_list=async_list,
    sync_list=sync_list,
)

1. Judge the Responses

   from trusteval import judge_responses
   target_models = ['your_target_model1', 'your_target_model2']
   judge_type = 'llm'  # Options: 'llm', 'vlm', 't2i'
   judge_key = 'your_judge_key'
   async_judge_model = ['your_async_model']
   await judge_responses(
       data_folder=base_dir,
       async_judge_model=async_judge_model,
       target_models=target_models,
       judge_type=judge_type,
   )

2. Generate Evaluation Metrics

   from trusteval import lm_metric
   lm_metric(
       base_dir=base_dir,
       aspect='ai_risk',
       model_list=target_models,
   )

3. Generate Final Report

   from trusteval import report_generator
   report_generator(
       base_dir=base_dir,
       aspect='ai_risk',
       model_list=target_models,
   )

Your report.html will be saved in the base_dir folder. For additional examples, check the examples folder.

A detailed trustworthiness evaluation report is generated for each dimension. The reports are presented as interactive web pages, which can be opened in a browser to explore the results. The report includes the following sections:

The data shown in the images below is simulated and does not reflect actual results.

Displays the evaluation scores for each model, with a breakdown of average scores across evaluation dimensions.

Summarizes the model's performance in the evaluated dimension using LLM-generated summaries, highlighting comparisons with other models.

Presents error cases for the evaluated dimension, including input/output examples and detailed judgments.

Shows the evaluation results for all models, along with visualized comparisons to previous versions (e.g., our v1.0 results).

We welcome contributions from the community! To contribute:

1. Fork the repository.
2. Create a feature branch (git checkout -b feature-name).
3. Commit your changes (git commit -m 'Add feature').
4. Push to your branch (git push origin feature-name).
5. Open a pull request.

@article{huang2025trustgen,
    title={On the Trustworthiness of Generative Foundation Models: Guideline, Assessment, and Perspective},
    author={Yue Huang and Chujie Gao and Siyuan Wu and Haoran Wang and Xiangqi Wang and Yujun Zhou and Yanbo Wang and Jiayi Ye and Jiawen Shi and Qihui Zhang and Yuan Li and Han Bao and Zhaoyi Liu and Tianrui Guan and Dongping Chen and Ruoxi Chen and Kehan Guo and Andy Zou and Bryan Hooi Kuen-Yew and Caiming Xiong and Elias Stengel-Eskin and Hongyang Zhang and Hongzhi Yin and Huan Zhang and Huaxiu Yao and Jaehong Yoon and Jieyu Zhang and Kai Shu and Kaijie Zhu and Ranjay Krishna and Swabha Swayamdipta and Taiwei Shi and Weijia Shi and Xiang Li and Yiwei Li and Yuexing Hao and Zhihao Jia and Zhize Li and Xiuying Chen and Zhengzhong Tu and Xiyang Hu and Tianyi Zhou and Jieyu Zhao and Lichao Sun and Furong Huang and Or Cohen Sasson and Prasanna Sattigeri and Anka Reuel and Max Lamparth and Yue Zhao and Nouha Dziri and Yu Su and Huan Sun and Heng Ji and Chaowei Xiao and Mohit Bansal and Nitesh V. Chawla and Jian Pei and Jianfeng Gao and Michael Backes and Philip S. Yu and Neil Zhenqiang Gong and Pin-Yu Chen and Bo Li and Xiangliang Zhang},
    journal={arXiv preprint arXiv:2502.14296},
    year={2025}
}

This project is licensed under the CC BY-NC 4.0.