-[2024/09] We are prototyping allowing users of LM Evaluation Harness to create and evaluate on text+image multimodal input, text output tasks, and have just added the `hf-multimodal` and `vllm-vlm` model types and `mmmu` task as a prototype feature. We welcome users to try out this in-progress feature and stress-test it for themselves, and suggest they check out [`lmms-eval`](https://github.com/EvolvingLMMs-Lab/lmms-eval), a wonderful project originally forking off of the lm-evaluation-harness, for a broader range of multimodal tasks, models, and features.
-[2024/07] [API model](docs/API_guide.md) support has been updated and refactored, introducing support for batched and async requests, and making it significantly easier to customize and use for your own purposes. **To run Llama 405B, we recommend using VLLM's OpenAI-compliant API to host the model, and use the `local-completions` model type to evaluate the model.**
-[2024/07] [API model](docs/API_guide.md) support has been updated and refactored, introducing support for batched and async requests, and making it significantly easier to customize and use for your own purposes. **To run Llama 405B, we recommend using VLLM's OpenAI-compliant API to host the model, and use the `local-completions` model type to evaluate the model.**
-[2024/07] New Open LLM Leaderboard tasks have been added ! You can find them under the [leaderboard](lm_eval/tasks/leaderboard/README.md) task group.
-[2024/07] New Open LLM Leaderboard tasks have been added ! You can find them under the [leaderboard](lm_eval/tasks/leaderboard/README.md) task group.
This method ensures that the right template is chosen based on the following:
0. If the model has no 'tokenizer' attribute: assumes that there is only a single possible chat template, handled on the model provider side internally. Returns the empty string.
1. If the model's tokenizer has multiple templates:
a. Use the specified template if it exists in the dictionary.
b. Use the default template from the list if no specific template is provided.
c. Raise an error if no default template exists and no specific template is provided.
2. If the model's tokenizer has a single template or no template:
a. Use the tokenizer's chat template if available.
b. Fall back to the default chat template if no tokenizer chat template exists.
Args:
chat_template (Union[bool, str]): Specifies the chat template to use.
- If False or None, no template is applied.
- If True, the default or only available template is used.
- If a string, the template with the matching name is used.
Returns:
Optional[str]: The selected chat template, or None if no template is applied.
"""
ifself.tokenizerisNone:
return""
ifchat_templateisFalseorchat_templateisNone:
eval_logger.warning(
"model.chat_template was called with the chat_template set to False or None. "
"Therefore no chat template will be applied. Make sure this is an intended behavior."
)
returnNone
# Convert boolean chat_template to None to ensure compatibility with the adapted logic
ifisinstance(chat_template,bool):
chat_template=None
using_default_template=False
# First, handle the cases when the model has a dict of multiple templates
f"Attempted to run tasks: {incompatible_tasks} which require multimodal input, but the selected model type does not currently implement this. Multimodal support is currently restricted to the ['hf-multimodal', 'vllm-vlm'] model type."
)
else:
raiseValueError(
f"Attempted to run tasks: {incompatible_tasks} which are text-only, but used a model type which only currently supports multimodal tasks."
| [aexams](aexams/README.md) | Tasks in Arabic related to various academic exams covering a range of subjects. | Arabic |
| [aexams](aexams/README.md) | Tasks in Arabic related to various academic exams covering a range of subjects. | Arabic |
| [agieval](agieval/README.md) | Tasks involving historical data or questions related to history and historical texts. | English, Chinese |
| [agieval](agieval/README.md) | Tasks involving historical data or questions related to history and historical texts. | English, Chinese |
| [anli](anli/README.md) | Adversarial natural language inference tasks designed to test model robustness. | English |
| [anli](anli/README.md) | Adversarial natural language inference tasks designed to test model robustness. | English |
| [arabic_leaderboard_complete](arabic_leaderboard_complete/README.md) | A full version of the tasks in the Open Arabic LLM Leaderboard, focusing on the evaluation of models that reflect the characteristics of Arabic language understanding and comprehension, culture, and heritage. Note that some of these tasks are machine-translated. | Arabic (Some MT) |
| [arabic_leaderboard_light](arabic_leaderboard_light/README.md) | A light version of the tasks in the Open Arabic LLM Leaderboard (i.e., 10% samples of the test set in the original benchmarks), focusing on the evaluation of models that reflect the characteristics of Arabic language understanding and comprehension, culture, and heritage. Note that some of these tasks are machine-translated. | Arabic (Some MT) |
| [arabicmmlu](arabicmmlu/README.md) | Localized Arabic version of MMLU with multiple-choice questions from 40 subjects. | Arabic |
| [arabicmmlu](arabicmmlu/README.md) | Localized Arabic version of MMLU with multiple-choice questions from 40 subjects. | Arabic |
| [arc](arc/README.md) | Tasks involving complex reasoning over a diverse set of questions. | English |
| [arc](arc/README.md) | Tasks involving complex reasoning over a diverse set of questions. | English |
| [arithmetic](arithmetic/README.md) | Tasks involving numerical computations and arithmetic reasoning. | English |
| [arithmetic](arithmetic/README.md) | Tasks involving numerical computations and arithmetic reasoning. | English |
"`group` and `group_alias` keys in tasks' configs will no longer be used in the next release of lm-eval. "
"`group` and `group_alias` keys in TaskConfigs are deprecated and will be removed in v0.4.5 of lm_eval. "
"`tag` will be used to allow to call a collection of tasks just like `group`. "
"The new `tag` field will be used to allow for a shortcut to a group of tasks one does not wish to aggregate metrics across. "
"`group` will be removed in order to not cause confusion with the new ConfigurableGroup "
"`group`s which aggregate across subtasks must be only defined in a separate group config file, "
"which will be the official way to create groups with addition of group-wide configurations."
"which will be the official way to create groups that support cross-task aggregation as in `mmlu`. "
"Please see the v0.4.4 patch notes and our documentation: https://github.com/EleutherAI/lm-evaluation-harness/blob/main/docs/new_task_guide.md#advanced-group-configs "
The Open Arabic LLM Leaderboard evaluates language models on a large number of different evaluation tasks that reflect the characteristics of the Arabic language and culture.
The benchmark uses several datasets, most of them translated to Arabic, and validated by native Arabic speakers. They also used benchmarks from other papers or prepared benchmarks from scratch natively for Arabic.
author = {Elfilali, Ali and Alobeidli, Hamza and Fourrier, Clémentine and Boussaha, Basma El Amel and Cojocaru, Ruxandra and Habib, Nathan and Hacid, Hakim},
abstract = "Recent advances in the space of Arabic large language models have opened up a wealth of potential practical applications. From optimal training strategies, large scale data acquisition and continuously increasing NLP resources, the Arabic LLM landscape has improved in a very short span of time, despite being plagued by training data scarcity and limited evaluation resources compared to English. In line with contributing towards this ever-growing field, we introduce AlGhafa, a new multiple-choice evaluation benchmark for Arabic LLMs. For showcasing purposes, we train a new suite of models, including a 14 billion parameter model, the largest monolingual Arabic decoder-only model to date. We use a collection of publicly available datasets, as well as a newly introduced HandMade dataset consisting of 8 billion tokens. Finally, we explore the quantitative and qualitative toxicity of several Arabic models, comparing our models to existing public Arabic LLMs.",
}
@misc{huang2023acegpt,
title={AceGPT, Localizing Large Language Models in Arabic},
author={Huang Huang and Fei Yu and Jianqing Zhu and Xuening Sun and Hao Cheng and Dingjie Song and Zhihong Chen and Abdulmohsen Alharthi and Bang An and Ziche Liu and Zhiyi Zhang and Junying Chen and Jianquan Li and Benyou Wang and Lian Zhang and Ruoyu Sun and Xiang Wan and Haizhou Li and Jinchao Xu},
year={2023},
eprint={2309.12053},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
@misc{lighteval,
author = {Fourrier, Clémentine and Habib, Nathan and Wolf, Thomas and Tunstall, Lewis},
title = {LightEval: A lightweight framework for LLM evaluation},
year = {2023},
version = {0.3.0},
url = {https://github.com/huggingface/lighteval}
}
```
### Groups and Tasks
*`arabic_leaderboard_alghafa`: A multiple-choice evaluation benchmark for zero- and few-shot evaluation of Arabic LLMs prepared from scratch natively for Arabic.
*`arabic_leaderboard_arabic_exams`: A question answering benchmark for high school examinations in different school subjects that requires knowledge and reasoning in different languages in multiple domains.
*`arabic_leaderboard_arabic_mmlu`: A multi-task language understanding benchmark for the Arabic language, sourced from school exams across diverse educational levels in different countries with native speakers in the region.
The data comprises multiple choice questions in 40 tasks.
*`arabic_leaderboard_arabic_mt_arc_challenge`: AI2 Reasoning Challenge (ARC) is a multiple-choice question task. The dataset contains only natural, grade-school science questions,
written for human tests. The challenge set contains only questions answered incorrectly by both a retrieval-based algorithm and a word co-occurence algorithm. (machine translated benchmark - part of the Alghafa Arabic translated LLM benchmark)
*`arabic_leaderboard_arabic_mt_boolq`: A true/false questions dataset that contains the columns passage, question, and the answer (i.e., true/false). (machine translated benchmark - part of the Alghafa Arabic translated LLM benchmark)
*`arabic_leaderboard_arabic_mt_copa`: Choice Of Plausible Alternatives (COPA) is a multiple-choice question dataset, which involves open-domain commonsense causal reasoning. (machine translated benchmark - part of the Alghafa Arabic translated LLM benchmark)
*`arabic_leaderboard_arabic_mt_hellaswag`: The tesk is to choose the next set of sentences, based on the given candidates. The tasks involve reading comprehension and information retrieval challenges
by testing the abilities of the models on basic knowledge (i.e., from 3rd grade to 9th) and commonsense inference. (machine translated benchmark - part of the Alghafa Arabic translated LLM benchmark)
*`arabic_leaderboard_arabic_mt_mmlu`: A multiple-choice question answering dataset from various branches of knowledge including humanities, social sciences, hard sciences, and other areas. The examples in the English dataset are translated into Arabic using ChatGPT with a translation prompt.
*`arabic_leaderboard_arabic_mt_openbook_qa`: A multiple-choice openbook question answering dataset that requires external knowledge and reasoning. The open book that comes with these questions is
based on elementary level science facts. (machine translated benchmark - part of the Alghafa Arabic translated LLM benchmark)
*`arabic_leaderboard_arabic_mt_piqa`: Physical Interaction Question Answering (PIQA) is a multiple-choice question answering based on physical commonsense reasoning. (machine translated benchmark - part of the Alghafa Arabic translated LLM benchmark)
*`arabic_leaderboard_arabic_mt_race`: A multiple-choice questions dataset to assess reading comprehension tasks based on English exams in China - designed for middle school and high school students
(machine translated benchmark - part of the Alghafa Arabic translated LLM benchmark)
*`arabic_leaderboard_arabic_mt_sciq`: A multiple-choice Science Question Answering task to assess understanding of scientific concepts about physics, chemistry, and biology. (machine translated benchmark - part of the Alghafa Arabic translated LLM benchmark)
*`arabic_leaderboard_arabic_mt_toxigen`: This benchmark consists of tasks designed to evaluate language models and classify input text as hateful or not hateful. (machine translated benchmark - part of the Alghafa Arabic translated LLM benchmark)
*`arabic_leaderboard_acva`: Arabic-Culture-Value-Alignment (ACVA) is a yes/no question dataset, generated by GPT3.5 Turbo from Arabic topics to assess model alignment with Arabic values and cultures.
For adding novel benchmarks/datasets to the library:
* [ ] Is the task an existing benchmark in the literature?
* [ ] Have you referenced the original paper that introduced the task?
* [ ] If yes, does the original paper provide a reference implementation? If so, have you checked against the reference implementation and documented how to run such a test?
If other tasks on this dataset are already supported:
* [ ] Is the "Main" variant of this task clearly denoted?
* [ ] Have you provided a short sentence in a README on what each new variant adds / evaluates?
* [ ] Have you noted which, if any, published evaluation setups are matched by this variant?
