Our client represents a government in the Arabian Gulf. In line with the state’s social development program, they needed a smart app that would provide residents and tourists with introductory information about the region’s sights, history, culture, and more, in natural language.
Most of our efforts were focused on developing language models. We needed to combine two language models and train both of them in English and Arabic, as existing open-source models did not meet the project’s specific requirements.
“The way the MetaDialog platform works for us is amazing; no other team could have trained such high-quality models in such a short time using our custom data. It answers exactly how we need it, according to our traditions, religion, rules, and regulations.”
Our platform is based on two powerful language models. The first LLM is able to generate text in both Arabic and English, surpassing GPT3.5 on six essential metrics and outperforming GPT4 on five out of six metrics. The second model serves as a cutting-edge textual retrieval program in Arabic, and is equally proficient in English. These models work together to enable the application to respond to a wide range of inquiries, including government services, cultural traditions, legal matters, religious topics, tourist attractions, and more.
The application is currently available to VIPs and citizens and is poised for future enhancements. These upgrades will introduce more advanced language models, enriching the application’s capabilities as it grows. The plan is to then scale the app to cover the entire country.
The MetaDialog development team delivered a custom, pre-trained Arabic-English language model, similar to the GPT model but improved and enhanced on several key metrics. We also introduced our own retrieval system for both languages, similar to BERT, which was modified according to product needs.