Overview
LMSYS and Kaggle have partnered to launch an exciting competition focused on human preference prediction, inviting participants to predict which responses users will favor in direct comparisons between Large Language Models (LLMs). This competition leverages a rich dataset sourced from the Chatbot Arena, featuring genuine conversations and user preferences across a diverse array of LLMs. By participating, you have the chance to enhance chatbot performance and ensure these models align more closely with user expectations. The training dataset is extensive, encompassing over 55,000 real-world dialogues and preferences, all with personally identifiable information rigorously removed to maintain privacy. Your submission will be evaluated against a hidden test set of 25,000 samples, ensuring a fair and challenging competition.
Background
Current benchmarks for LLMs often do not capture the nuances of real-world usage, leading to a gap between how models perform in tests and the satisfaction users derive from their interactions. Tools like the Chatbot Arena allow users to pose questions and express preferences for different responses, yet this valuable data has not been fully utilized to build models that predict and optimize for user preferences at scale. Accurately predicting these preferences is vital for creating conversational AI that resonates with human users and fosters an engaging experience. Winning models could enable LLMs to adapt their responses dynamically, tailoring interactions to individual user preferences across various scenarios. Furthermore, this competition seeks to identify the underlying factors influencing user choices beyond mere correctness; many user inquiries are open-ended, revealing a correlation between user preference and qualities such as conversational tone and engagement. This competition also presents an excellent opportunity for testing reward modeling strategies within Reinforcement Learning from Human Feedback (RLHF) algorithms.
Competition Details
This thrilling competition is set to run until August 5th, offering a total prize pool of $100,000. This includes a grand prize of $25,000 for the top entrant, alongside $20,000 rewards for those finishing in 2nd to 4th place, and $15,000 for the 5th place winner. This is a unique opportunity to contribute to the advancement of human-aligned language models while also gaining profound insights into human preferences and decision-making processes. These findings could hold significant implications for both the fields of computer science and psychology, illuminating the factors that shape human preferences in conversational AI.
Inspired by: Source
