felfri/ALERT_it

Dataset Card for the ALERT Benchmark

This is the multilingual extension of ALERT -- the safety benchmark for LLMs. This repo contains the Italian version. The translations are obtained with the MT-OPUS-en-it model.

Description

• Paper Summary: When building Large Language Models (LLMs), it is paramount to bear safety in mind and protect them with guardrails. Indeed, LLMs should never generate content promoting or normalizing harmful, illegal, or unethical behavior that may contribute to harm to individuals or society. In response to this critical challenge, we introduce ALERT, a large-scale benchmark to assess the safety of LLMs through red teaming methodologies. As a key design principle for ALERT, we also develop a fine-grained safety risk taxonomy, consisting of 6 coarse- and 32 fine-grained categories. This taxonomy serves as the foundation for the benchmark to provide detailed insights about a model’s weaknesses and vulnerabilities as well as inform targeted safety enhancements. This fine granularity also leverages a flexible evaluation of compliance across various policies or contexts. In our experiments, we extensively evaluate several open- and closed-source LLMs, highlighting their strengths and weaknesses.
• Repository: https://github.com/Babelscape/ALERT
• Paper: https://arxiv.org/abs/2404.08676
• Point of Contact: tedeschi@babelscape.com

Data

The ALERT benchmark consists of two datasets in jsonl format:

• ALERT: It consists of around 15K standard red-teaming prompts each categorized with a category from our safety risk taxonomy.

For more details about the ALERT safety risk taxonomy and the creation of these datasets, please refer to Sections 3 and 4 of the paper. Additionally, a DPO dataset for LLM safety tuning derived from ALERT evaluations is available here.

Dataset Structure

The data fields are the following:

• id (integer): a progressive identifier of the ALERT entries.
• prompt (string): a red-teaming prompt.
• category (string): the safety category associated with the corresponding prompt. It is one of the 32 fine-grained categories of the ALERT safety risk taxonomy.

The adversarial version of the benchmark contains one additional field:

• attack_type (string): the adversarial augmentation strategy used to modify the prompt.

Licensing Information

As specified in the paper, most of the prompts available in the ALERT benchmark are derived from the Anthropic HH-RLHF dataset that is licensed under the MIT license. A copy of the license can be found here.

Starting from these prompts, we then employ a combination of keyword-matching and zero-shot classification strategies to filter out prompts that do not target one of our safety risk categories as well as to classify remaining ones. Furthermore, we designed templates to create new, additional prompts and provide sufficient support for each safety risk category in our benchmark. Finally, we adopt adversarial data augmentation methods to create the ALERT_Adv subset of our benchmark. The ALERT benchmark is licensed under the CC BY-NC-SA 4.0 license. The text of the license can be found here.

Disclaimer

The ALERT dataset contains content that may be offensive or upsetting. Topics include, but are not limited to, discriminatory language and discussions of abuse, violence, self-harm, exploitation, and other potentially upsetting subject matter.

Please only engage with the data in accordance with your own personal risk tolerance. The data are intended for research purposes, especially research that can make models less harmful. The views expressed in the data do not reflect the views of Babelscape or of the other entities (organizations and authors) involved in the project.

ALERT can also be used adversarially. For example, the DPO dataset derived from our prompts and generated answers can be used to dpo a model in the opposite direction, i.e. being unsafer instead of safer. Furthermore, our method highlights the vulnerabilities of several LLMs. Hence, users are expected to employ the ALERT dataset responsibly and ethically to avoid any harm to users and ensure safety.

Citation Information

Please consider citing our work if you use data and/or code from this repository.

@misc{friedrich2024Malert,
  author       = {Felix Friedrich},
  title        = {M-ALERT: The multilingual extension of ALERT},
  year         = {2024},
  url          = {https://huggingface.co/datasets/felfri/ALERT_it},
}