Test-time Augmentation for Factual Probing

Environment

• GPU: CUDA 11.6
• python3.9 (miniconda4-4.7.12)

Dataset

This dataset consists of 12,500 unique facts from wikidata. Each fact has a subject, a relation, and an object. Facts are sampled with truncated sampling to reduce the biased distribution of the objects. The familiarity of a subject/object is controlled by the number of related facts in wikidata. It is also a disjoint set of LAMA trex

Generate with TTA

python script/main.py