Advancing Generative AI: How Hugo Boosted Meta’s Llama-2

The AI landscape is fiercely competitive, with leading tech companies racing to develop advanced generative AI tools. Amid this urgency, Meta faced the challenge of rapidly enhancing its Generative AI capabilities to accelerate the deployment of its Large Language Models (LLMs). The goal was ambitious: process vast amounts of data and provide superior responses across a broad spectrum of subjects, from STEM to humanities and everyday interactions.

In a landmark project, Meta developed Llama 2, an open-source LLM designed to increase access and create economic and social opportunities for individuals, developers, researchers, creators, and businesses of all sizes. However, advancing such models requires responsible innovation and reliable outsourcing partners.

Hugo, a leader in customer support and data annotation—was one of five vendors carefully selected for breadth of experience, quality output, and ability to scale efficiently to meet the project’s requirements. “Choosing Hugo for this project was pivotal; we knew their expertise in precision annotation and commitment to scalability would lay the foundation for groundbreaking advancements in generative AI technology,” explains Meta’s LLM Core Project Manager.

This case study highlights Hugo’s key contributions to a project that pushed the boundaries of AI and showcased the power of effective outsourcing partnerships.

Pioneering AI Advancements with Hugo’s Expertise

Meta had a challenging goal: to deploy their LLM in a production environment as swiftly as possible. This required teams capable of processing massive volumes of data, tackling complex technical challenges, and feeding their models with superior responses to prompts.

At Hugo, commitment to excellence knows no bounds, and with Llama 2, our teams raised the bar by outperforming other vendors by 25% and scaling the team 10x within weeks. Over a few months, Hugo’s team had fine-tuned prompts for the LLM across 150+ subjects, guided by four core principles.

1. Team Construction

With a keen eye for talent, Hugo assembled a team of 600+ subject matter experts for Llama 2. Each specialist was selected for domain expertise, proven experience with complex annotations, and attention to detail, ensuring alignment with Meta’s objectives.

2. Training & Development

Prior to production, annotators underwent an intensive 80-hour training program that simulated real-world scenarios, equipping them to handle complex tasks from day one. This training focused on critical areas such as entity recognition, sentiment analysis, and intent classification to enhance the model’s understanding.

3. Quality Assurance

A dual-layer quality assurance process was implemented. The first layer focused on context alignment, ensuring that annotated data accurately reflected the intended meaning. The second layer emphasized factual accuracy, ensuring that the information was not only correct but also coherent and fluent.

4. Continuous Workflow Optimization

Hugo utilized its expertise in experimental design to optimize workflows and generate actionable insights for process improvement. The learning and development team, which included QA managers and internal consultants, performed thorough analyses, such as:

• Assessing annotation accuracy to pinpoint areas for improvement.
• Reviewing annotator feedback to enhance training and refine processes.

Daily targeted strategies were implemented for continuous improvement, including:

• Refining training materials based on performance data.
• Adjusting workflows to address common challenges during retraining.

“Hugo’s consistent quality and reliability in managing complex AI projects have been unparalleled,” the Project Manager recounts. “They have become an integral part of our success story in AI development.”

Key Achievements

• Achieved a data accuracy rate of 95%, exceeding the client’s SLA of 80% and outperforming other vendors, who had a performance rate of 70%.
• Outperformed other vendors by 25%.
• Scaled the team 10x in just a few weeks.
• Fine-tuned prompts across 150+ subjects.
• Improved response accuracy from 87% to 95% in 8 weeks.
• Increased the model’s ability to handle diverse topics by 40%.
• Demonstrated flexibility and scalability to meet project demands without compromising output quality, earning recognition as the leading vendor.
• Hugo’s QA rubric was adopted by Meta’s engineering teams for use in other generative AI projects and recognized as the Gold Standard by other vendors.

A Testament to Value-Driven Partnerships

Hugo’s partnership with Meta on the Llama 2 project highlights a shared vision for AI innovation and responsible development. By prioritizing ethical AI practices, Hugo ensured that data privacy and bias reduction measures were integral to every stage of the process. The Project Manager states, “Hugo’s involvement demonstrates their commitment to delivering inclusive outcomes and effectively tackling complex projects.”

This value-driven partnership reinforces Hugo’s reputation as a reliable, innovative, and quality-driven outsourcing partner. As a result, Meta has expanded its collaboration with Hugo to encompass other advanced AI initiatives, reflecting confidence in Hugo’s ability to meet complex requirements with precision and scalability.