Prosus fine-tunes open-source LLMs on Databricks to deploy GenAI across its portfolio companies

Overview

Prosus, a global technology investor with an applied AI lab of over 600 data scientists, used Databricks to pretrain and fine-tune open-source foundation models like MPT-30B on data such as Stack Overflow content, building conversational GenAI solutions including a Polish-language marketplace model and a Portuguese-language model for a food-delivery investee company. Databricks let Prosus productionize LLMs faster and more cost-effectively across its portfolio without building its own GenAI training infrastructure.

The challenge

Prosus's AI team found that infrastructure complexity and lack of processing power limited its ability to innovate with LLMs, and many cloud platforms on the market were not optimized to train and run its in-house LLMs, while building its own compute clusters was often unreliable and highly resource-intensive to build and maintain.

The solution

Prosus turned to Databricks as a fully managed platform to test and train models at scale, starting with the value of Stack Overflow data on open source Databricks foundation models using the MPT-30B LLM for pretraining and fine-tuning, then using a broader, more diverse dataset to run smaller models, identify points of failure, and retrain, refine and tune models to meet business requirements including conversational chatbot outputs.

Reported business value

Databricks let Prosus productionize models faster and at a lower cost, enabling it to deliver AI solutions like a Polish-language online marketplace model and a Portuguese-language food-delivery model to non-English-speaking users across its portfolio companies.