Building compliant-ready generative AI solutions
Dynamo leverages Databricks to train secure multilingual foundation models
Days to pretrain an 8-billion
parameter multilingual LLM
Speedup in model training
Dynamo AI builds multilingual foundation models that democratize access to compliant and secure GenAI. Recently, they used Databricks Mosaic AI Training to train their 8-billion parameter multilingual LLM in just 10 days.
Dynamo AI was founded in 2022 to empower companies to safely adopt AI solutions in a rapidly changing landscape of emerging regulations and compliance challenges. Their product suite spans the entire AI development lifecycle, ensuring the secure and responsible production and deployment of enterprise-grade AI systems. Eric Lin, Head of Applied AI at Dynamo AI, explains, “We enable enterprises to train, develop and then deploy models in a way that’s compliant with safety regulations. We take a lot of the cutting-edge research we learned from our academic roots at MIT and Harvard and deploy it at scale for millions of customers.”
Addressing the security risks of GenAI adoption
Dynamo AI’s products focus on three different phases of the generative AI model lifecycle: evaluation, remediation and deployment. As Lin stated, “Our customers see the amazing potential with GenAI models. However, they face challenges in making sure that the model is providing the right information and is not harming either their business or the consumer. What we do is build guardrails and evaluation suites around these models to make sure they can foster this trust in the real world.”
In order to support their customers’ needs, Dynamo AI knew their platform would be required to incorporate the latest research on privacy and security while providing guardrails against attacks and vulnerabilities. They also wanted to be able to implement a range of pre-built policies based on specific industries or regions, as well as support multilingual tasks. “We wanted a foundation model that can do specific kinds of enterprise-compliant guardrailing, and we didn’t see any open source model that could do the same thing.” Dynamo AI decided to train its own foundation model, Dynamo 8B, and turned to the team at Databricks Mosaic AI for help.
Training a safe, robust and reliable foundation model
Dynamo AI used Databricks Mosaic AI Training to build a foundation model that would support their AI evaluation, remediation and deployment workloads. Compared to their early experimentation on other platforms, the Dynamo AI team found immediate benefits to working with Databricks. “Your out-of-the-box training scripts were really useful and helped us save weeks’ worth of development time in setting up training runs,” explained Lin. “For example, we were experimenting with different architectures. Originally, our foundation model didn’t have the efficiency gains that we expected. But with help from the Databricks Mosaic Research team to address the efficiency issues, we were able to train the model around 20% faster than what we found on competitors.” In fact, with Mosaic AI Training’s built-in speedups and GPU availability on demand, it took Dynamo AI just 10 days to pretrain an 8-billion parameter multilingual LLM.
Lin also appreciated the hands-on expertise that the Databricks team brought to the project. “We had an issue during our training run where we were using an unexpected amount of memory, which slowed us down because we had to lower the batch size. Your team was able to look at our model weights and help us figure out where memory leakage was happening. Once we solved that issue, our training went much faster.”
Dynamo AI relied on the Databricks Mosaic AI team’s experience in other ways. Before training their Dynamo 8B LLM, it was necessary to gather and prepare vast amounts of data. Explained Lin, “We were prioritizing multilingual data because we wanted to make sure our model could serve many different countries, so we did a lot of public dataset aggregation.” Dynamo AI’s team was able to estimate how much data would be needed to reach their performance benchmarks based on previous work done by Databricks researchers in training their MPT and DBRX open source models.
From model training to pre-built industry solutions
Thanks to Databricks, Dynamo AI has gone from model experimentation to revenue generation in just a few months. Today, Dynamo AI offers a range of solutions powered by their multilingual foundation models. These product modules enable compliance, privacy and security throughout the generative AI stack, in support of sensitive workloads such as multilingual customer support, customer onboarding, claims processing and fraud detection. Many of their customers are also leveraging the data and AI capabilities of the Databricks Data Intelligence Platform, which makes it easy for them to securely power Dynamo AI’s GenAI applications within their secure Databricks environment.
Although the Dynamo AI team was able to train their Dynamo 8B foundation model in just a few weeks, Mosaic AI Training continues to be a tool critical to their success. Explained Lin, “We are continually integrating research into our LLM to better defend against new vulnerabilities. Our goal is to build an application layer for security and compliance that provides the safeguards organizations need to thrive in the age of AI.” To learn more about Dynamo AI’s suite of privacy, security and compliance solutions, visit www.dynamo.ai.