search

Mivg / SLED

The official repository for Efficient Long-Text Understanding Using Short-Text Models (Ivgi et al., 2022) paper
MIT License
67 stars 9 forks source link

generated_output in usage_example.py doesn't contain a valid output #5

Closed mrcabbage972 closed 1 year ago

mrcabbage972 commented 1 year ago

In usage_example.py, generated_output is always equal to the prefix. Or if there's no prefix, it's equal to the document content. I tried it with a few different texts.

Mivg commented 1 year ago

Hi, The model loaded in usage_example.py is the pertained facebook/bart-base model wrapped with SLED. As it was not fine-tuned on anything, it indeed mostly just generates output identical to the input as it was trained to Denise the input. The usage example just gives an example how one can load and generate using a model, you should fine-tune on your task before using it. Alternatively, you can load a fine-tuned SLED checkpoint from the hub. For example, you can try loading BART-base-SLED fine-tuned on contract NLI with tau/bart-base-sled-contractnli