Task1578 and Task288 about natural-instructions

[danyaljj]

Do we have many other duplicate tasks?

I think do, we just haven't found them yet :)

Although in this particular case what I meant was the explanations for positive and negative examples are better in task 288.

I think we're optimizing two objectives: more diversity for our annotations (more examples, more definitions, etc.) while reducing noise and redundancy (fewer duplications, fewer garbage examples, etc.) So, if the examples/explanations are not quite duplicate/noisy (as in, a human can read and be, at least partially, convinced), I'd say let's retain them.

[Palipoor]

Task 288 has 6500 instances and task 1578 has almost 1900. Both are probably randomly sampled from Gigaword. Also, task 288 has better explanations. I guess we can keep 288 and drop 1578.
Also, task 1579 is "incorrect summary generation" and task 289 is verifying whether a summary is correct. I think task 289 makes more sense. An incorrect summary can be anything.
@danyaljj Should I drop those two?

[danyaljj]

Here is a suggestion:

• I think there is value in having multiple task definitions. In one PR, we can change the "Definition": "" field to "Definition": [""].
• In followup PRs, we can merge the tasks that are equivalent and based on the same underlying dataset.

What do you think?

[Palipoor]

It sounds good to me! Although in this particular case what I meant was the explanations for positive and negative examples are better in task 288. Do we have many other duplicate tasks?

[danyaljj]

@Palipoor wondering if you can help with merging these tasks?

[Palipoor]

Sure! Will do tomorrow.