The BELLS project
BELLS (Benchmarks for the Evaluation of LLM Safeguards) s a project aimed at evaluating the reliability and generality of supervision systems (input-output safeguard) for large language models (LLMs). These systems are responsible for analyzing the inputs and outputs of LLMs while they operate, in order to detect critical behaviors: hallucinations in their response, attempt to jailbreak by the user, or even mobilization of dangerous capacity (hacking, etc.).
Many products are available to meet some of these needs, but there is no methodology to evaluate them yet.
The BELLS project aims to establish an independent assessment system, to meet three critical needs:
- Evaluate and compare existing supervision tools (such as Lakera Guard, Azure Prompt Flow, or NVIDIA NeMo Guardrails)
- Stimulate the development of tools capable of detecting yet unknown anomalies in order to be better prepared for future emerging problems
- Encourage the creation of systems suitable for a wide variety of LLM applications
We collaborate with a diversity of stakeholders to ensure the industrial and academic relevance of our work. We are currently working with Vincent Corruble, Associate Professor at Sorbonne University, and Giskard, a startup specializing in AI application testing.
For more details, you can check:
Main missions
1. Contribute to the development of the BELLS project:
- Design and implement new evaluation tests for the systems that monitor LLMs
- Develop results visualization and analysis tools
- Participate in the creation of a red-teaming platform for supervision systems
- Ensure scientific monitoring of advances in the supervision of LLMs and propose new evaluation methodologies
- Maintain and improve the BELLS project software infrastructure to enable modular evaluation of various supervision systems
2. Ensure the operational implementation of the BELLS project
- Initiate and facilitate interactions with different project stakeholders (academics, companies, volunteer open-source contributions, etc.)
- Prise d'initiative dans la direction technique du projet (infrastructure, format des tests, etc.)
3. Possibly support other technical projects from CeSIA, particularly on demos and prototypes
Desired profile
Technical skills:
- Very proficient in Python
- Mastery of software development best practices
- Experience with LLM application development
- Knowledge of the safety issues of LLMs
- Familiarity with data visualization tools (Streamlit, etc.)
Qualities:
- Autonomy, initiative and leadership: we are looking for a manager able to lead the project even in the absence of detailed directions from the hierarchy.
- Strong capacity for analysis and synthesis
- Excellent communication skills (English required, French desirable)
- Ability to interact with a wide diversity of stakeholders (industry, academia, public institutions, etc.)
- Ability to manage a small team (potentially including volunteers and interns)
About CeSIA
CeSIA is an initiative launched and incubated by EffiSciences, which aims to promote responsible development of AI through awareness-raising actions, R&D, and support for actors of the AI ecosystem.
Sample of our track record:
- Our first publication on the BELLS project was accepted at the NextGen AI Safety workshop at ICML 2024. We also presented the project at the conference AI_Dev organized by the Linux Foundation.
- More than 1,000 students have participated in our events and training, more than 30 of whom have gone on to direct their careers toward AI security.
- Teaching of 4 accredited courses at ENS: two programs on AI safety; a biosecurity program; a program on engaged research
Working conditions
- Hours: Full-time.
- Offices: We use a coworking space located in Arcueil (5 minutes' walk from the Laplace stop), which is available to you from Monday to Friday, but with no obligation to use it.
How to apply
To apply, please fill out this form before August 11th end of day.
We will assess the applications on a rolling basis. The next step will be a 30-minute interview. If you have any questions or concerns, please do not hesitate to contact us at florent@securite-ia.fr
