The client, based in the United States, is a global leader in technology and innovation, renowned for its groundbreaking advancements in artificial intelligence (AI) and machine learning (ML).

The client sought the highest-quality proprietary datasets trained by effective SFT and RLHF techniques to enhance its largest and most general AI model. The complexity and challenging nature of their datasets required experienced LLM trainers to improve reasoning, coding, and other high-level cognitive capabilities. Given the scope of their model and the fast-paced nature of the tech industry, a flexible solution that could adapt to frequently changing guidelines while maintaining high-quality data standards was necessary.

The client, in collaboration with Turing, initiated a meticulous model enhancement strategy, focusing on Implicit Code Execution (ICE) and code reasoning. Turing deployed an on-demand team of LLM advisors and trainers for code-related RLHF, data clean-up, and SFT prompt engineering, including:

• Notebook process: The team developed a process for SFT data analysis and generating single- and multi-turn notebooks. Single-turn (ST) notebooks considered each input in isolation, while multi-turn (MT) notebooks maintained context.
• Notebook rectification: The team rectified pre-generated notebooks and iterated updates based on new training insights.
• Feedback and data curation: The team provided feedback on the client's RLHF tool, curated data from PDFs and .xlsx files, labeled log data, and performed code-related RLHF involving rewrites.

Through effective collaboration and efficient workflows, the client now enjoys an LLM capable of enhanced scientific problem-solving, coding and debugging, and more. Turing’s contributions include:

• Notebook generation: Over 50,000 high-quality ST and MT notebooks, and more, generated—significantly enhancing the model's training data.
• Data accuracy: Rectified pre-generated notebooks, improving proprietary data accuracy and reliability.
• Curated evaluation datasets: Single-turn evaluation datasets were curated from PDFs, .xlsx files, and complex file types, ensuring comprehensive model assessment.