Vijil provides private cloud services that harden AI models during development, defend models during operation, and evaluate reliability, security, and safety continuously.

Vijil

Vijil is an AI software startup on a mission to help developers build and operate intelligent agents that people can trust. Many organizations today are drawn to the productive potential of agents but are held back from production use because large language models have inherent vulnerabilities to attack and propensities for harm. For AI developers at enterprises and startups that want to measure and mitigate these risks, Vijil offers cloud services that continuously improve the trustworthiness of custom models based on open-source LLMs.

vijil

"Testing and mitigating toxicity — inherent in LLMs trained on Internet data — is important" 

If you had asked me a month ago about how to measure (much less mitigate) toxicity in LLMs, I wouldn't have had a good answer. The parameters/size of training data in the most popular models presents a prohibitively expensive pruning process. OctoAI partnered with vijil to evaluate WizardLM-8x22B, Mixtral-8x22B (base model), and Mixtral 8x7B Instruct. Vijil tested these models for several categories of toxicity. The Vijil evaluation framework obtains test prompts from the realtoxicityprompts (RTP) benchmark dataset, generated prompts from a custom adversarial model, and hand-crafted by red-team experts. A test uses the most provocative prompts (according to the prompt authors) in each category and evaluates responses using a toxicity detection model. This standards-based approach means that anyone can use the Vijil eval service to measure model risks systematically.

Have you used a toxicity mitigation technique on LLMs before or is there more you'd be interested in learning? Link in comments for the blog post.

We evaluated Meta Llama3 and Llama 2 for the propensity to perpetuate harmful stereotypes in response to jailbreaking prompts. Using 10K diverse prompts, we tested if the model agrees with common stereotypes associating negative attributes that might lead to unfair outcomes in the workplace for people in EEOC protected groups. Turns out, there are stark differences between Llama 3 and Llama 2, and within the demographic groups. Leave us a comment or suggestion. https://lnkd.in/gW4J-6DU

When OctoAI asked Vijil if we could test the toxicity of WizardLM-2-8x22B, we kicked off our tests just as the topic came up in the meeting and our tests were done shortly afterwards. We went on to compare these results to those of base Mixtral-8x22B and Mixtral-8x7B-Instruct just as quickly. If you want to deploy models in production on OctoAI, Vijil can shorten your time-to-trust. Reach out and book a demo!

Over the past 6 months, we’ve hosted several CXO Insight Calls centered around #AI and seen that there’s a real need to solve a number of core issues before #IT leaders can move the technology to production in the enterprise. According to Forbes, 55% of business leaders say their teams remain resistant to adopting AI. Hear from leaders at startups Securiti, Sema4.ai and vijil on how to leverage external technology to drive practical advances in AI today:

🚨 Webinar Alert 🚨

Vijil is happy to announce our first webinar, to be held next Thursday, 25 January 2024, at 10-11 am US Pacific time!

With new AI security vulnerabilities popping up left and right, there is a need for structured testing of LLMs and agents for known vulnerabilities. The open-source project garak: LLM vulnerability scanner allows you to do just that. In this upcoming webinar, we are glad to host garak team members: founder and lead maintainer Dr Leon Derczynski, who is Principal Research Scientist at NVIDIA, and active contributor Subho Majumdar, PhD, our Head of AI.

If you are interested in finding and reporting LLM vulnerabilities at scale, join us in learning about garak from Leon and Subho.

Space is limited, so register today using the link below!
https://lnkd.in/gRtqiGjU

We're #hiring an Applied Scientist to focus on the research and development of metrics and mechanisms for #trustworthyAI. Apply today or share this post with an ML researcher you know.

AI:VIJIL

Vijil

Analysis Not Available