FM 2 DS : Few-Shot Multimodal Multihop Data Synthesis with Knowledge Distillation for Question Answering
Amirhossein Abaskohi, Spandana Gella, Giuseppe Carenini, Issam H. Laradji · 2024
Read FM 2 DS : Few-Shot Multimodal Multihop Data Synthesis with Knowledge Distillation for Question Answering by Amirhossein Abaskohi, Spandana Gella, Giuseppe Carenini, Issam H. Laradji (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-07030.txt.
Description
Amirhossein Abaskohi, Spandana Gella, Giuseppe Carenini, and Issam Laradji present, in a 2024 paper, a factory for building hard question-answering data. The target skill is answering questions that need several hops across both images and text. Good training sets for that skill barely exist, since most sets cover one hop, one medium, or short texts.
Their framework, FM2DS, runs in five stages. It gathers linked documents from Wikipedia, drafts high-level questions and answers with a teacher model, and filters them hard for quality. Models trained on the made-up data beat models trained on human-gathered data by about two points on two benchmarks. The team also release a thousand-item test set for long documents, polished by human checkers.
Catalog record
| CREATOR | Amirhossein Abaskohi, Spandana Gella, Giuseppe Carenini, Issam H. Laradji |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 75 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-07030