Contrastive Learning for Fake News
Collection
Datasets and model for training models with contrastive loss. Domain/task: Fake news / hoax classification.
•
15 items
•
Updated
This dataset infers a contradiction case between facts and contents from HOAX class subset in nlp-brin-id/id-hoax-report-merge-v2.
The subsets can be utilized as samples for interleaving batch sampling during training stage of contrastive learning modles.
Attributes used = 'Content', 'Fact'.
See 'Files and Versions' for inspecting the subset independently: