NVIDIA Dominates KDD Cup 2024 Data Science Competition

Team NVIDIA’s Clean Sweep: Team NVIDIA, consisting of six NVIDIANs, secured first place across all five competition tracks at the prestigious Amazon KDD Cup 2024, demonstrating their mastery in generative AI and data science:

• The team’s innovative approach involved generating 500,000 questions using a combination of manual creation, large language models, and transforming existing e-commerce datasets to overcome the limited training data provided by the organizers.
• By fine-tuning the Qwen2-72B model using eight NVIDIA A100 Tensor Core GPUs and employing the QLoRA technique, Team NVIDIA outperformed all competitors despite the constraints imposed by the competition’s format.

KDD Cup 2024: Mimicking Online Shopping Complexities with LLMs: The KDD Cup 2024, organized by ACM SIGKDD and hosted by Amazon, focused on leveraging large language models to enhance the online shopping experience:

• The challenge aimed to test participants’ models using the ShopBench benchmark, which replicates the complexities of online shopping with 57 tasks and approximately 20,000 questions derived from real-world Amazon shopping data.
• The competition evaluated models across four key shopping skills: shopping concept understanding, shopping knowledge reasoning, user behavior alignment, and multilingual abilities, as well as an all-in-one challenge that combined all tasks from the previous tracks.

NVIDIA’s Winning Solution: Team NVIDIA’s winning approach involved creating a single model for each track, utilizing a combination of data generation, fine-tuning, and quantization techniques:

• The team generated training datasets based on the provided examples and synthesized additional data using Llama 3 70B hosted on build.nvidia.com.
• They fine-tuned the Qwen2-72B model using eight NVIDIA A100 Tensor Core GPUs for approximately 24 hours, employing the QLoRA training process to efficiently modify a subset of the model’s weights.
• The model was quantized with AWQ 4-bit and used the vLLM inference library to predict the test datasets on four NVIDIA T4 Tensor Core GPUs within the time constraints.

Broader Implications: Team NVIDIA’s victory at the KDD Cup 2024 highlights the growing importance of generative AI and data science in solving complex, real-world challenges:

• The team’s success in mimicking the intricacies of online shopping using LLMs demonstrates the potential for AI to revolutionize e-commerce and enhance the customer experience.
• NVIDIA’s clean sweep for the second consecutive year underscores the company’s leadership in AI and its commitment to pushing the boundaries of what is possible with generative AI and data science.
• The team’s planned submission of a detailed paper on their solution and presentation at KDD 2024 in Barcelona will further contribute to the advancement of AI research and its real-world applications.