Spaces:
Running
on
CPU Upgrade
Running
on
CPU Upgrade
CHANGELOG_TEXT = f""" | |
## [2023-06-13] | |
- Adjust description for TruthfulQA | |
## [2023-06-12] | |
- Add Human & GPT-4 Evaluations | |
## [2023-06-05] | |
- Increase concurrent thread count to 40 | |
- Search models on ENTER | |
## [2023-06-02] | |
- Add a typeahead search bar | |
- Use webhooks to automatically spawn a new Space when someone opens a PR | |
- Start recording `submitted_time` for eval requests | |
- Limit column max-width | |
## [2023-05-30] | |
- Add a citation button | |
- Simplify Gradio layout | |
## [2023-05-29] | |
- Auto-restart every hour for the latest results | |
- Sync with the internal version (minor style changes) | |
## [2023-05-24] | |
- Add a baseline that has 25.0 for all values | |
- Add CHANGELOG | |
## [2023-05-23] | |
- Fix a CSS issue that made the leaderboard hard to read in dark mode | |
## [2023-05-22] | |
- Display a success/error message after submitting evaluation requests | |
- Reject duplicate submission | |
- Do not display results that have incomplete results | |
- Display different queues for jobs that are RUNNING, PENDING, FINISHED status | |
## [2023-05-15] | |
- Fix a typo: from "TruthQA" to " | |
QA" | |
## [2023-05-10] | |
- Fix a bug that prevented auto-refresh | |
## [2023-05-10] | |
- Release the leaderboard to public | |
""" | |
TITLE = """<h1 align="center" id="space-title">π€ Open LLM Leaderboard</h1>""" | |
INTRODUCTION_TEXT = f""" | |
π With the plethora of large language models (LLMs) and chatbots being released week upon week, often with grandiose claims of their performance, it can be hard to filter out the genuine progress that is being made by the open-source community and which model is the current state of the art. The π€ Open LLM Leaderboard aims to track, rank and evaluate LLMs and chatbots as they are released. | |
π€ A key advantage of this leaderboard is that anyone from the community can submit a model for automated evaluation on the π€ GPU cluster, as long as it is a π€ Transformers model with weights on the Hub. We also support evaluation of models with delta-weights for non-commercial licensed models, such as LLaMa. | |
π In the **first tab (LLM Benchmarks)**, we evaluate models on 4 key benchmarks from the <a href="https://github.com/EleutherAI/lm-evaluation-harness" target="_blank"> Eleuther AI Language Model Evaluation Harness </a>, a unified framework to test generative language models on a large number of different evaluation tasks. In the **second tab (Human & GPT Evaluations)**, the evaluations are performed by having humans and GPT-4 compare completions from a set of popular open-source language models (LLMs) on a secret set of instruction prompts. | |
""" | |
LLM_BENCHMARKS_TEXT = f""" | |
Evaluation is performed against 4 popular benchmarks: | |
- <a href="https://arxiv.org/abs/1803.05457" target="_blank"> AI2 Reasoning Challenge </a> (25-shot) - a set of grade-school science questions. | |
- <a href="https://arxiv.org/abs/1905.07830" target="_blank"> HellaSwag </a> (10-shot) - a test of commonsense inference, which is easy for humans (~95%) but challenging for SOTA models. | |
- <a href="https://arxiv.org/abs/2009.03300" target="_blank"> MMLU </a> (5-shot) - a test to measure a text model's multitask accuracy. The test covers 57 tasks including elementary mathematics, US history, computer science, law, and more. | |
- <a href="https://arxiv.org/abs/2109.07958" target="_blank"> TruthfulQA </a> (0-shot) - a test to measure a modelβs propensity to reproduce falsehoods commonly found online. | |
We chose these benchmarks as they test a variety of reasoning and general knowledge across a wide variety of fields in 0-shot and few-shot settings. | |
""" | |
HUMAN_GPT_EVAL_TEXT = f""" | |
Evaluation is performed by having humans and GPT-4 compare completions from a set of popular open-source language models (LLMs) on a secret set of instruction prompts. The prompts cover tasks such as brainstorming, creative generation, commonsense reasoning, open question answering, summarization, and code generation. Comparisons are made by humans and a model on a 1-8 Likert scale, where the labeler is required to choose a preference each time. Using these preferences, we create bootstrapped Elo rankings. | |
We collaborated with **Scale AI** to generate the completions using a professional data labeling workforce on their platform, [following the labeling instructions found here](https://docs.google.com/document/d/1c5-96Lj-UH4lzKjLvJ_MRQaVMjtoEXTYA4dvoAYVCHc/edit?usp=sharing). To understand the evaluation of popular models, we also had GPT-4 label the completions using this prompt. | |
For more information on the calibration and initiation of these measurements, please refer to the [announcement blog post](https://huggingface.co/blog/llm-leaderboard). We would like to express our gratitude to **LMSYS** for providing a [useful notebook](https://colab.research.google.com/drive/1lAQ9cKVErXI1rEYq7hTKNaCQ5Q8TzrI5?usp=sharing) for computing Elo estimates and plots. | |
""" | |
EVALUATION_QUEUE_TEXT = f""" | |
# Evaluation Queue for the π€ Open LLM Leaderboard, these models will be automatically evaluated on the π€ cluster | |
""" | |
CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results" | |
CITATION_BUTTON_TEXT = r"""@misc{open-llm-leaderboard, | |
author = {Edward Beeching, Sheon Han, Nathan Lambert, Nazneen Rajani, Omar Sanseviero, Lewis Tunstall, Thomas Wolf}, | |
title = {Open LLM Leaderboard}, | |
year = {2023}, | |
publisher = {Hugging Face}, | |
howpublished = "\url{https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard}" | |
} | |
@software{eval-harness, | |
author = {Gao, Leo and | |
Tow, Jonathan and | |
Biderman, Stella and | |
Black, Sid and | |
DiPofi, Anthony and | |
Foster, Charles and | |
Golding, Laurence and | |
Hsu, Jeffrey and | |
McDonell, Kyle and | |
Muennighoff, Niklas and | |
Phang, Jason and | |
Reynolds, Laria and | |
Tang, Eric and | |
Thite, Anish and | |
Wang, Ben and | |
Wang, Kevin and | |
Zou, Andy}, | |
title = {A framework for few-shot language model evaluation}, | |
month = sep, | |
year = 2021, | |
publisher = {Zenodo}, | |
version = {v0.0.1}, | |
doi = {10.5281/zenodo.5371628}, | |
url = {https://doi.org/10.5281/zenodo.5371628} | |
} | |
@misc{clark2018think, | |
title={Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge}, | |
author={Peter Clark and Isaac Cowhey and Oren Etzioni and Tushar Khot and Ashish Sabharwal and Carissa Schoenick and Oyvind Tafjord}, | |
year={2018}, | |
eprint={1803.05457}, | |
archivePrefix={arXiv}, | |
primaryClass={cs.AI} | |
} | |
@misc{zellers2019hellaswag, | |
title={HellaSwag: Can a Machine Really Finish Your Sentence?}, | |
author={Rowan Zellers and Ari Holtzman and Yonatan Bisk and Ali Farhadi and Yejin Choi}, | |
year={2019}, | |
eprint={1905.07830}, | |
archivePrefix={arXiv}, | |
primaryClass={cs.CL} | |
} | |
@misc{hendrycks2021measuring, | |
title={Measuring Massive Multitask Language Understanding}, | |
author={Dan Hendrycks and Collin Burns and Steven Basart and Andy Zou and Mantas Mazeika and Dawn Song and Jacob Steinhardt}, | |
year={2021}, | |
eprint={2009.03300}, | |
archivePrefix={arXiv}, | |
primaryClass={cs.CY} | |
} | |
@misc{lin2022truthfulqa, | |
title={TruthfulQA: Measuring How Models Mimic Human Falsehoods}, | |
author={Stephanie Lin and Jacob Hilton and Owain Evans}, | |
year={2022}, | |
eprint={2109.07958}, | |
archivePrefix={arXiv}, | |
primaryClass={cs.CL} | |
}""" | |
VISUALIZATION_TITLE = """<h1 align="center" id="space-title">π Visualizations</h1>""" | |
PLOT_1_TITLE = "Fraction of Model A Wins for All Non-tied A vs. B Comparisons" | |
PLOT_2_TITLE = "Comparison Count of Each Combination of Models (not allowing ties)" | |
PLOT_3_TITLE = "Elo Estimates with error bars (ties allowed)" | |
PLOT_4_TITLE = "Fraction of Model A Wins for All Non-tied A vs. B Comparisons" | |