Open research toolkit from the ICML 2024 TrustLLM benchmark for evaluating LLM trustworthiness across six dimensions via CLI and Python.
TrustLLM is a research toolkit for measuring how trustworthy large language models are across six dimensions: truthfulness, safety, fairness, robustness, privacy, and ethics. It packages the ICML 2024 benchmark so you can run the datasets against local Hugging Face weights or any OpenAI-compatible API, then keep the data, settings, and outputs together in one run directory.
The workflow is split into generation and scoring. The generate step collects model responses per task, with checkpoints, bounded retries, and an HTML report of completion counts. The evaluate step then scores a complete set of responses, using rules, a downloaded classifier, embeddings, or an API judge depending on the task. The base install is lightweight, and optional extras add local model loading and scoring dependencies.
The project also documents CLI and Python orchestration with JSON results, so the same pipeline can be driven by scripts or AI agents.
Install the base package from GitHub (requires Git)
python -m pip install "trustllm @ git+https://github.com/HowieHwong/TrustLLM.git@main#subdirectory=trustllm_pkg"Download the benchmark dataset
python -m trustllm download --output dataSet your API key and endpoint, then run a smoke-test generation
export OPENAI_API_KEY="your-api-key"
export OPENAI_BASE_URL="https://your-provider.example/v1"
python -m trustllm generate \
--backend api --model your-model-id \
--task safety --data data/dataset \
--limit 5 --concurrency 4 --output runs/api-safetySummary generated by AI from the project's README. GitHub stats refresh daily.
Python library that scores LLM responses for likely hallucinations using uncertainty quantification, with black-box, white-box, judge and ensemble scorers.
Python library that generates and runs robustness, bias, fairness and accuracy tests against NLP models and LLMs with one line of code.
An evaluation harness for testing LLMs as autonomous agents, with 300 human-verified tasks graded on completion, safety, and robustness.
A Python CLI that audits AI agents against 61 inspections, testing whether they stay within their role, resist manipulation and avoid fabrication.
An open-source command-line scanner that probes LLMs for failure modes such as prompt injection, data leakage, jailbreaks, and toxic output.
An open-source Python framework for building robust RL environments that run agent tasks in isolated VMs to train aligned AI.