Learn More with Less Education: Smart Compression Method with Unspecified Guide

With impressive results to solve complex problems of large language models, these capabilities are achieved only with more data and calculation strength. A new research shows that this mold can be broken.

The study published in arXiv suggests a framework called the stage narrowing** with the guidance of injustice. This method makes it possible to seriously reduce the size of supervisory data in model training:

• Focus on examples that the model is most uncertain, instead of the entire educational data
• Providing large-scale decline in training time and cost
• Gain productivity increase by limiting performance loss minimum
• Gradually becoming more selective in the model with staged approach
• Redesigning the requirements of Supervision in the Post-training phase

The main question of the research is quite simple but important: how much can the logic audit data be compressed? The available methods are spent on tremendous resources to train with larger data sets, which shows the work done is strategic selection small data sets can reach the same goal more efficiently.

This approach can be transformative for organizations and researchers, especially limited to resources. Having high-performance models with low-cost training can make artificial intelligence technology accessible to a wider audience.