5 ms·
This might provide some guidance: http://gltr.io/ http://gltr.io/
by mrdrozdov 4y ago
This might provide some guidance: http://gltr.io/ http://gltr.io/
- Der_Einzige 4y agoThis and related techniques are trivially foolable by fine-tuning the model. They're also trivially foolable by using sampling techniques or settings which encourage the model to generate rare words a lot. Also foolable with filter-assisted decoding: https://paperswithcode.com/paper/most-language-models-can-be-poets-too-an-ai https://paperswithcode.com/paper/most-language-models-can-be...