PaLI-3 architecture에
pre-training recipe : consists of two backbones, ViT-2B and TextEncoder UL2-3B
Pre-training: Chart2Table Mixture
unfrozen ViT에 대해 수행한다.
여러 Chart to Table 데이터 mixture을 사용해 pretraining을 수행한다.
Fine-tuning: Multi-task Loss
two ways of incorporating the rationales available in the extended datatset.
1. Single-Task setup
changing the target task from answer to rationale, answer
2. Multi-Task setup
answer and rationale are treated as independent task
Result
Singletask vs. Multitask
Human dataset에 비해 Augmented dataset이 QA pair이 좀 더 단조로움
Learning with augmented dataset
'공부기록 > [Paper]' 카테고리의 다른 글
| Time-LLM: Time Series Forecasting by Reprogramming Large Language Models(240414, 공부기록 2일차) (2) | 2024.04.13 |
|---|
