Multi-Query Convergence: A Strategy for Boosting Language Model Performance
Minjun Kim, Yeongseo Ha, Jaechang Shim · Journal of Korea Multimedia Society · 2024
This investigation implemented a multi-query convergence approach on Meta’s 라마 3.1 language model, significantly enhancing its capabilities on the GLUE dataset tasks, specifically WNLI and MRPC. By adopting this methodology, the 라마 3.1 Queried model showed marked improvements, with accuracy in the MRPC task rising from 50% to 62%. In the WNLI task, accuracy also improved significantly, increasing from 39.43% to 61.97%. These enhancements are largely due to the method’s effectiveness in enhancing complex contextual understanding and detailed language pattern recognition. Further evaluations measured additional performance metrics, notably the Certainty Ratio and Time Per Prompt. The Certainty Ratio for the MRPC task was 98.21% with the original model and 93.83% with the queried model, while for the WNLI task, it was 100.00% originally and 98.32% with the queried model. This slight decrease in certainty illustrates a nuanced trade-off in model confidence when handling more complex queries. Additionally, the Time Per Prompt metric underscored the efficiency of the queried model, maintaining competitive processing times despite the added complexity of the multi-query approach. These metrics not only validate the model’s enhanced accuracy but also its practical applicability and reliability in processing language tasks efficiently.