akhmat-s/

t5-large-quant-grammar-corrector

FlanT5 was trained using JFLEG dataset. The primary objective of the experiment was to develop a highly effective tool using relatively small models, minimal datasets, and constrained computational resources.

To accomplish this goal, we implemented two key strategies:

• Perplexity-Based Data Pruning With Small Reference Models.
• A simple sampling and voting method for multiple LLM agents. model was trained.