Comparison of mT5-large and Flan-T5-large Inference Efficiency in Zero-Shot Cross-Lingual Transfer on XTREME Benchmarks

Intermediate-task training---fine-tuning a pretrained model on an intermediate task before fine-tuning again on the target task---often improves model performance substantially on language understanding tasks in monolingual English settings. We investigate whether English intermediate-task training is still helpful on non-English target tasks. Using nine intermediate language-understanding tasks, we evaluate intermediate-task transfer in a zero-shot cross-lingual setting on the XTREME benchmark. We see large improvements from intermediate training on the BUCC and Tatoeba sentence retrieval tas

Research goal: How does the inference efficiency (tokens per second) of mT5-large compare to Flan-T5-large when performing zero-shot cross-lingual transfer on XTREME benchmarks after intermediate-task training, while maintaining equivalent accuracy?

Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 9.2/10.