Initialize components with interfaces and stub implementations #5
Conversation
viswavi
changed the title
There was a problem hiding this comment.
Thanks @viswavi ! This is looking good.
I left some comments on only the dataset generators part, but many of them are general philosophical comments that should apply throughout the codebase. Please take a look. If you agree with the comments you can reflect them, and if you don't understand or disagree with anything please follow up and we can discuss!
|
|
||
| """ | ||
| _ = prompt_spec, num_examples, split # suppress unused variable warnings | ||
| return datasets.Dataset.from_pandas(pd.DataFrame({})) |
There was a problem hiding this comment.
In general, it seems like a reasonable restriction that the datasets should be the same length as num_examples. This example here breaks that restriction, so maybe we should make it a dataset of appropriate length full of empty dictionaries instead?
There was a problem hiding this comment.
Done. @zhaochenyang20 when we write unit tests for dataset_generator we should add a check that the dataset is of the same length as num_examples. I've created an issue (#7 ) to add this.
There was a problem hiding this comment.
I will try to create our unit test for DatasetGenerator this weekend following our new restriction. 🙂
Co-authored-by: Graham Neubig <neubig@gmail.com>
Co-authored-by: Graham Neubig <neubig@gmail.com>
Co-authored-by: Graham Neubig <neubig@gmail.com>
Co-authored-by: Graham Neubig <neubig@gmail.com>
Co-authored-by: Graham Neubig <neubig@gmail.com>
Description
Initialize all Prompt2Model components with separate interfaces and stub implementations. This PR also contains an integration test (using all the components' stub implementations) that verifies that the types of all the components fit together.
References
Blocked by