NVIDIA's AI Masters Dominate KDD Cup 2024 with Groundbreaking Wins
In a remarkable display of technological prowess, NVIDIA's AI Masters have won first place across all five competition tracks in the prestigious KDD Cup 2024, according to the NVIDIA Blog. The annual data science competition, hosted by Amazon and organized by the Association for Computing Machinery’s Special Interest Group on Knowledge Discovery and Data Mining (ACM SIGKDD), saw NVIDIA's team excelling in every category.
Competition Overview
The KDD Cup 2024 focused on the theme “Multi-Task Online Shopping Challenge for LLMs,” requiring participants to solve complex challenges using limited datasets. Competitors were tasked with addressing issues in text generation, multiple-choice questions, named entity recognition, ranking, and retrieval. The competition aimed to push the boundaries of large language models (LLMs) by providing only 96 example questions, insufficient to train a model traditionally.
Chris Deotte, a senior data scientist at NVIDIA, explained, “The new trend in LLM competitions is that they don’t give you training data. They give you 96 example questions — not enough to train a model — so we came up with 500,000 questions on our own.”
Team NVIDIA, consisting of Ahmet Erdem, Benedikt Schifferer, Chris Deotte, Gilberto Titericz, Ivan Sorokin, and Simon Jegou, generated a variety of questions by writing some themselves, using a large language model to create others, and transforming existing e-commerce datasets. This innovative approach allowed them to fine-tune a language model effectively.
NVIDIA’s Winning Strategy
The NVIDIA team utilized the newly released Qwen2-72B, a large language model with 72 billion parameters. They fine-tuned this model using eight NVIDIA A100 Tensor Core GPUs, employing a technique known as Quantized Low-Rank Adaptation (QLoRA) for efficient training. The GPUs significantly reduced the time required for fine-tuning, allowing the team to process data quickly and effectively.
First, the team created training datasets based on the provided examples and synthesized additional data using Llama 3 70B hosted on build.nvidia.com. They then employed QLoRA, which modifies a smaller subset of the model’s weights, allowing for efficient training and fine-tuning. The model was quantized using 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.
This comprehensive approach secured NVIDIA top positions in each individual track, culminating in an overall first-place finish—a clean sweep for the second consecutive year.
About the KDD Cup 2024
The KDD Cup is an esteemed annual competition that promotes research and development in data science. This year's challenge, hosted by Amazon, aimed to replicate the complexities of online shopping, using the ShopBench benchmark. The benchmark included 57 tasks and about 20,000 questions derived from real-world Amazon shopping data, focusing on four key shopping skills and a fifth “all-in-one” challenge:
- Shopping Concept Understanding: Decoding complex shopping concepts and terminologies.
- Shopping Knowledge Reasoning: Making informed decisions with shopping knowledge.
- User Behavior Alignment: Understanding dynamic customer behavior.
- Multilingual Abilities: Shopping across languages.
- All-Around: Solving all tasks from the previous tracks in a unified solution.
NVIDIA’s innovative approach and cutting-edge technology have once again proven their dominance in the field of AI and data science. The team plans to submit a detailed paper on their solution next month and present their findings at KDD 2024 in Barcelona.