Is there a way to obtain the summaries generated by abstractor during each train step

Hi developer,

Based on your model, the abstractor should be able to generate summaries, then together with ground truth summaries, the reward is calculated and fed to extractor.

So how could we get these generated summaries in each train step?

I do appreciate!

No file, test, or entry point is named. Start by locating the training step where the abstractor produces summaries and where rewards are calculated before being fed to the extractor. Done means establishing how those per-step generated summaries can be obtained, with the behavior verified in the training flow.

Written by the indexing model from the issue text.