Hyperscalers OCR Evaluation

[samiuc]

• Implements new API encapsulation layer for dataset creation and prediction providers based on this PR
• Evaluation dataset is a subset of original pixparse-idl-wds
• Note: versions of torch and docling are temporarily pinned. (I was facing issues installing 2.5.1 torch version locally) - will be reverted prior to merge.

[cau-git]

@samiuc Thanks for this update. I recognize some of the comments on the previous PR are now addressed, without checking deeply.

To make sure this will integrate cleanly, I set the base branch to cau/new-class-design as it should be. However after doing so, the diff shows that you took units from this branch and re-defined them (potentially with changes) in the main source tree docling_eval with different module paths. I must therefore ask you to ensure that all the work in this PR is cleanly building on top of cau/new-class-design, with the goal that it merges back into that branch. As it is, it won't. Are there any particular reasons why you would touch the DatasetRecord class or put your dataset builders somewhere else than where cau/new-class-design suggests?

[cau-git]

@samiuc I went through the changes, and have a few remarks below.
The main point I see needs to be addressed is that there must not be a dataset-specific PredictionProvider for pixparse. The API design decouples prediction providers and dataset builders, they must both work independently from one another. Can you please decompose the code for that?

Additionally, I would like you to check what is the overlap of your implementation for Azure DI with the work of @praveenmidde on this PR, which also implements a PredictionProvider for Azure within the table dataset evaluation. Eventually, we should have only one prediction provider for Azure DI.

[cau-git]

@samiuc I decomposed this PR into three new ones with updates. Let's please continue the work there, then close this PR.

• feat: PixParse OCR dataset builder #61
• feat: AWS Textract and Google DocAI Prediction providers #62
• feat: OCR evaluator #63

[cau-git]

Closing this since it is superseded by the newer PRs.