Flexible crowdsourcing platform powers large-scale model industry with high-quality data and efficient human alignment

On August 23, Dr. Wu Runze, technical director of NetEase Fuxi User Portrait Group, was invited to participate in the large model industry theme forum with the theme of "Boiling Capital, AGI Riding the Waves". At the forum, he gave a speech on the theme of "Efficient Human Alignment for Large Model Implementation Applications". He introduced to the relevant enterprises in the large model industry present how NetEase Fuxi helps create a closed loop of large model data, and shared how to build a large model data closed loop at a low cost. High-quality data-based cases and experiences.

Among the three elements of large models - data, computing power, and algorithms, strengthening the scale of pre-training models and improving data quality are key methods to achieve better artificial intelligence effects. However, simply increasing the model size does not necessarily lead to better results. In the context of subjectivity present in many real-world tasks, scaling up the model can lead to unreliable and questionable results. Therefore, in order to ensure the credibility and effectiveness of artificial intelligence, a comprehensive optimization strategy that takes into account data, computing power, and algorithms is needed.

In response to the above situation, Dr. Wu Runze said in his speech that NetEase Fuxi, as the first domestic game artificial intelligence research institution laboratory, can use the massive data and simulation environment of the game platform to achieve To promote the development of artificial intelligence technology. After NetEase Fuxi was selected into Zhejiang Province's "Pioneer Project" relying on the "Ultra-large-scale Pre-training Cloud Platform" project, it tried to accumulate and explore technology from multiple aspects such as data, algorithms, systems, and applications. When faced with the problem of increasing the capabilities of large models, It was decided to conduct directional guidance through human alignment, introduce positive, well-intentioned human feedback into the large model, and introduce "crowd consensus" as a feedback signal.

Currently, NetEase Fuxi Youling crowdsourcing platform is committed to providing high-quality data solutions for generative AI. Relying on NetEase's rich experience and deep integration in the fields of game engines and AI, we actively assist large model manufacturers to solve the closed-loop problems of large-scale computing power data and pre-training models. Our goal is to help the industry obtain higher quality data at a lower cost, thereby promoting the healthy development of the industry.

The above is the detailed content of Flexible crowdsourcing platform powers large-scale model industry with high-quality data and efficient human alignment. For more information, please follow other related articles on the PHP Chinese website!