GitLab

* fix incorrect lookback protections

* bump generate_until task versions

* multimedqa

* Update medqa.yaml

* move to benchmarks folder

* add README.md

---------
Co-authored-by: Lintang Sutawika <lintang@sutawika.com>

* Refine scoring logic for multiple_target "exact_match" metric

* skip old tests from master

* skip old tests from master

* delete tests from master

Over a dozen papers have used the updated citation block, but Google Scholar has noticed none of them. Since it does understand this citation, I think we should use it going forward until we have a way to ensure the newer citations are actually logged.

* do not ensure ascii

* Update __main__.py

---------
Co-authored-by: Lintang Sutawika <lintang@sutawika.com>

* add hellaswag_nl

* add other languages and update readme to hellaswag

* refactor as new task

* update readme

* add endline to yaml files and readme.md

* add group, change folder location and update yaml file

* rename default hellaswag yaml file

* fix whitespace error in some labels

* downgrade log level of whitespace checking

---------
Co-authored-by: JorgeDeCorte <jorge.decorte@ravago.be>
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

* Remove self.dataset_path post_init process

* Update task.py

* Update task.py

* copies max_length from huggingface

* handle max_length properly

* get tokens from inputs

* substitute Collator for Reorderer

* `batch=auto` if using data_parallel

* nit

* cleanup

* update code comments

* `ray.shutdown()` after calling method if data_parallel_size > 1

---------
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

* auto-batch requires len of iter

* handle case when batch_size="auto:N"

* Add example failing task

This task includes an invalid import. This will cause an exception and
the task will not be loaded. But this just results in a DEBUG level log
message, so in normal usage you'll see no error, and will be told the
task doesn't exist.

Here's an example command line to run the task:

    python -m lm_eval --model hf --model_args pretrained=rinna/japanese-gpt-1b --tasks fail

This task is based on a Japanese Winograd task, but that's not
important, and was just used due to familiarity.

* Do not ignore errors when loading tasks

* Change how task errors are logged

This makes the proposed changes from PR discussion.

1. Exceptions not related to missing modules/imports are logged as
   warnings.

2. module/import related exceptions are still logged at debug level, but
   if any of them happen there is a warning about it with instructions
   on how to show logs.

* Remove intentionally failing task

---------
Co-authored-by: Paul O'Leary McCann <polm@dampfkraft.com>

* fix group

* siqa: default.yml -> default.yaml

* max_gen_toks -> self.max_gen_toks

* add ids to task tests

* fix siqa

* fix gen_kwargs for openai-chat

* Add remove_whitespace to FLD benchmark

* bump task version

---------
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

* refactor dataloader

* cleanup + add docs

* change arg

* renamed Collator and added testing

* parametrized test for Collator

* appease pre-commit

* added edge case batch 0 (no batching)

* fix typos

---------
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

* modularize HFLM code

* pass through extra kwargs to AutoModel.from_pretrained call

* remove explicit model_kwargs

* rename gptq -> autogptq

* fix tokenizer pad token errors

* ensure model always respects device_map and autogptq's selected devices

* add a _get_config helper fn

* add mambaLMWrapper

* add mamba extra

* add mamba extra

* fix conditional import

* Fix botched merge commit

* Remove beginning-of-file comment for consistency

* Add docstring for mambaLM re: supported kwargs

* Alphabetize extras

* Update extras table

* appease precommit

* run precommit on mamba_lm

* Add retry error handler

* fixup! Add retry error handler

* Move to utils.py

* Run isort on utils.py

* Catch multiple exceptions

* Update LMs with exception handler

* Fixes to anthropic retry handler

* fix callback kwarg

* Update textsynth.py

* fix python 3.8 incompatibility

* fix indenterror I introduced

* placate linter?

* Update on_exception_callback kwarg name

* fixup! Merge branch 'main' into add-retry-error-handler

* fixup! fixup! Merge branch 'main' into add-retry-error-handler

* Merge conflicts are fun

* Run pre-commit

---------
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

* Update README.md

Add a not about running on apple arm gpus

* Update README.md

* Update README.md

---------
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>

* remove tokenizer for openai chat completions

* reordering function

* linter

* remove tiktoken import

* change version field formatting in metadata

* mention versioning in new task guide

* add instructions for changelog

* run linters

* separate local flag

* tokenizer_backend

* import order

* LocalChatCompletionsLM add

* clean up completions class

* clean up completions class

* update tokens

* README

* fix constructor

* eos token

* folding local-chat-completions into OpenAIChatCompletions

* refactoring to include gen_kwargs as passable option

* add todo on chat completion kwarg validation

* Ruff and README fix

* generalize to **kwargs

* remove unnecessary kwargs

* README and remove kwargs

* README

* update kmmlu default formatting

* Update _default_kmmlu_yaml

* Delete lm_eval/tasks/kmmlu/utils.py