Groundcocoa

[HarshKohli]

Adding the groundcocoa benchmark for evaluating the conditional & compositional reasoning of language models in a grounding task

[CLAassistant]

All committers have signed the CLA.

[baberabb]

Hi! Thanks for the PR. Couple of issues:

1. Test failing because the context ends with \n and the target delimiter is " ". So the full sample will be for e.g:

<context>

<space>The answer is A

I think we can set target_delimiter="" in the config?
2. Could you also delete the test_output folder