GitLab

* add xcopa dataset

* add xstory_cloze dataset and run pre-commit

* fix xcopa validation and test sets

* add xwinograd dataset

* add pawsx task

* add xnli task

* update task table with recently added tasks

* remove unused metrics from paws-x

* add mgsm task and fix gsm8k

* fix gsm8k until

* update task table

* Support bigbench-hard json tasks using multiple_choice_grade

* Add support for greedy decoding in bigbench tasks

* move bigbench_resources to datasets

* rectify changes to rf.greedy_until w upstream

* make path to resource import reflect new location

---------
Co-authored-by: haileyschoelkopf <hailey.schoelkopf@yale.edu>

ll_rolling computes adaptive bs separately and passes computed bs to _ll_tokens to avoid recomputation

This reverts commit 5fe7e2c0, reversing
changes made to 14b8ec7e.

- saves the doc idx by order in dataset
- saves the full ctx. we can turn this off it is too much memory
- saves the idx of the fewshot examples
- saves the source (test/val) of the fewshot examples
- TODO: We now need to return the pred/target from process_results
- TODO: Make it so override-ed process-resutls don't have to re-impl
this.

Updated doc: If the answer choices is empty, then it is generation; else ranked choice. This will be the canonical approach when using PS.

Turned off generation tasks for now. Changed process to look at the metrics. Only accuracy implemented.