GitLab

* allow for hf-causal to take dtype arg

* document this change

* 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

When `limit` is <1, limit represents the percentage of the total number of examples.
If it is >=1,  then it means the number of examples per task (only use this for testing).

Fix automatic batching for greedy_until
Fix stop sequences bug for greedy_until
Fix max length bug when model max length is not specified in tokenizer

* 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>

* Add 8 bit support

* added trust remote code

* Update main.py

* Update huggingface.py

* Update huggingface.py

* Fix bugs

* Fix PR #422 issues

* fix(style): remove empty newline

* fix(stlye,args): run `pre-commit` and re-order args for backward-compat

* fix(style): pre-commit the `README.md`

* host wikitext on hf

* transfer wikitext to Eleuther HF org

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