Stability of Zero-Shot Cross-Lingual Transfer on XTREME-R with Intermediate-Task Fine-Tuning

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 stability of zero-shot cross-lingual transfer on XTREME-R vary when using different intermediate-task datasets (e.g., NLI, QA, STS) in English-only fine-tuning versus multilingual fine-tuning, measured by accuracy variance across target languages?

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