File size: 12,349 Bytes
6ed68b6
d52179b
 
 
 
460d762
 
 
 
5601a63
 
 
0227006
 
 
b3f0642
 
 
 
b29b985
 
 
 
460d762
b29b985
2a73469
 
 
 
46f8d78
2a73469
58733e4
46f8d78
6ed68b6
2a73469
 
6ed68b6
 
2a73469
6ed68b6
 
2a73469
 
 
 
6ed68b6
 
a7cba30
6ed68b6
 
2a73469
6ed68b6
 
2a73469
6ed68b6
58733e4
 
 
 
6e8f400
58733e4
a0b557b
d2e8eca
d16cee2
a66fcca
 
0227006
 
 
d2e8eca
6e8f400
 
 
 
58733e4
 
 
18916e3
58733e4
217b585
58733e4
d2e8eca
d16cee2
 
256c5d3
d16cee2
 
 
 
 
 
 
 
 
 
 
256c5d3
edd3aac
d16cee2
256c5d3
d16cee2
 
256c5d3
d16cee2
 
256c5d3
 
 
 
 
 
 
 
 
788108a
d2e8eca
 
 
 
 
 
 
d16cee2
 
 
 
 
12cea14
 
 
 
 
b323764
 
 
 
 
 
 
d16cee2
 
 
 
 
58733e4
 
 
6e8f400
 
58733e4
2a73469
 
217b585
 
d2e8eca
2a73469
 
 
d06dc21
e61a555
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6e8f400
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
CHANGELOG_TEXT = f"""
## [2023-06-19]
- Added model type column
- Hid revision and 8bit columns since all models are the same atm

## [2023-06-16]
- Refactored code base
- Added new columns: number of parameters, hub likes, license

## [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 AutoEvalColumn 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">πŸ€— Open LLM Leaderboard</h1>"""

INTRODUCTION_TEXT = f"""
πŸ“ The πŸ€— Open LLM Leaderboard aims to track, rank and evaluate LLMs and chatbots as they are released. 

πŸ€— Anyone from the community can submit a model for automated evaluation on the πŸ€— GPU cluster, as long as it is a πŸ€— Transformers model with weights on the Hub. We also support evaluation of models with delta-weights for non-commercial licensed models, such as the original LLaMa release.

Other cool benchmarks for LLMs are developped at HuggingFace, go check them out: πŸ™‹πŸ€– [human and GPT4 evals](https://huggingface.co/spaces/HuggingFaceH4/human_eval_llm_leaderboard), πŸ–₯️ [performance benchmarks](https://huggingface.co/spaces/optimum/llm-perf-leaderboard)

🟒: Base pretrained model – πŸ”Ά: Instruction finetuned model – 🟦: Model finetuned with RL  (read more details in "About" tab)
"""

LLM_BENCHMARKS_TEXT = f"""
# Context
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. 

πŸ“ˆ 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. 

- <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 model’s propensity to reproduce falsehoods commonly found online. Note: TruthfulQA in the Harness is actually a minima a 6-shots task, as it is prepended by 6 examples systematically, even when launched using 0 for the number of few-shot examples.

For all these evaluations, a higher score is a better score. 
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.

# Some good practices before submitting a model

### 1) Make sure you can load your model and tokenizer using AutoClasses:
```python
from transformers import AutoConfig, AutoModel, AutoTokenizer
config = AutoConfig.from_pretrained("your model name", revision=revision)
model = AutoModel.from_pretrained("your model name", revision=revision)
tokenizer = AutoTokenizer.from_pretrained("your model name", revision=revision)
```
If this step fails, follow the error messages to debug your model before submitting it. It's likely your model has been improperly uploaded. 

Note: make sure your model is public!
Note: if your model needs `use_remote_code=True`, we do not support this option yet but we are working on adding it, stay posted!

### 2) Convert your model weights to [safetensors](https://huggingface.co/docs/safetensors/index)
It's a new format for storing weights which is safer and faster to load and use. It will also allow us to add the number of parameters of your model to the `Extended Viewer`!

### 3) Make sure your model has an open license!
This is a leaderboard for Open LLMs, and we'd love for as many people as possible to know they can use your model πŸ€—

### 4) Fill up your model card
When we add extra information about models to the leaderboard, it will be automatically taken from the model card

# Reproducibility and details

### Details and logs
You can find:
- detailed numerical results in the `results` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/results
- details on the input/outputs for the models in the `details` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/details
- community queries and running status in the `requests` Hugging Face dataset: https://huggingface.co/datasets/open-llm-leaderboard/requests

### Reproducibility
To reproduce our results, here is the commands you can run, using [this version](https://github.com/EleutherAI/lm-evaluation-harness/tree/b281b0921b636bc36ad05c0b0b0763bd6dd43463) of the Eleuther AI Harness:
`python main.py --model=hf-causal --model_args="pretrained=<your_model>,use_accelerate=True,revision=<your_model_revision>"`
` --tasks=<task_list> --num_fewshot=<n_few_shot> --batch_size=2 --output_path=<output_path>`

The total batch size we get for models which fit on one A100 node is 16 (8 GPUs * 2). If you don't use parallelism, adapt your batch size to fit. 
*You can expect results to vary slightly for different batch sizes because of padding.*

The tasks and few shots parameters are:
- ARC: 25-shot, *arc-challenge* (`acc_norm`)
- HellaSwag: 10-shot, *hellaswag* (`acc_norm`)
- TruthfulQA: 0-shot, *truthfulqa-mc* (`mc2`)
- MMLU: 5-shot, *hendrycksTest-abstract_algebra,hendrycksTest-anatomy,hendrycksTest-astronomy,hendrycksTest-business_ethics,hendrycksTest-clinical_knowledge,hendrycksTest-college_biology,hendrycksTest-college_chemistry,hendrycksTest-college_computer_science,hendrycksTest-college_mathematics,hendrycksTest-college_medicine,hendrycksTest-college_physics,hendrycksTest-computer_security,hendrycksTest-conceptual_physics,hendrycksTest-econometrics,hendrycksTest-electrical_engineering,hendrycksTest-elementary_mathematics,hendrycksTest-formal_logic,hendrycksTest-global_facts,hendrycksTest-high_school_biology,hendrycksTest-high_school_chemistry,hendrycksTest-high_school_computer_science,hendrycksTest-high_school_european_history,hendrycksTest-high_school_geography,hendrycksTest-high_school_government_and_politics,hendrycksTest-high_school_macroeconomics,hendrycksTest-high_school_mathematics,hendrycksTest-high_school_microeconomics,hendrycksTest-high_school_physics,hendrycksTest-high_school_psychology,hendrycksTest-high_school_statistics,hendrycksTest-high_school_us_history,hendrycksTest-high_school_world_history,hendrycksTest-human_aging,hendrycksTest-human_sexuality,hendrycksTest-international_law,hendrycksTest-jurisprudence,hendrycksTest-logical_fallacies,hendrycksTest-machine_learning,hendrycksTest-management,hendrycksTest-marketing,hendrycksTest-medical_genetics,hendrycksTest-miscellaneous,hendrycksTest-moral_disputes,hendrycksTest-moral_scenarios,hendrycksTest-nutrition,hendrycksTest-philosophy,hendrycksTest-prehistory,hendrycksTest-professional_accounting,hendrycksTest-professional_law,hendrycksTest-professional_medicine,hendrycksTest-professional_psychology,hendrycksTest-public_relations,hendrycksTest-security_studies,hendrycksTest-sociology,hendrycksTest-us_foreign_policy,hendrycksTest-virology,hendrycksTest-world_religions* (`acc` of `all`)

### Quantization
To get more information about quantization, see:
- 8 bits: [blog post](https://huggingface.co/blog/hf-bitsandbytes-integration), [paper](https://arxiv.org/abs/2208.07339)
- 4 bits: [blog post](https://huggingface.co/blog/4bit-transformers-bitsandbytes), [paper](https://arxiv.org/abs/2305.14314)

### Icons
🟒 means that the model is pretrained
πŸ”Ά that it is finetuned
🟦 that is was trained with RL.
If there is no icon, we have not uploaded the information on the model yet, feel free to open an issue with the model information!


# In case of model failure
If your model is displayed in the `FAILED` category, its execution stopped. 
Make sure you have followed the above steps first. 
If everything is done, check you can launch the EleutherAIHarness on your model locally, using the above command without modifications (you can add `--limit` to limit the number of examples per task).

"""

EVALUATION_QUEUE_TEXT = f"""
# Evaluation Queue for the πŸ€— Open LLM Leaderboard
These models will be automatically evaluated on the πŸ€— cluster.
"""

CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
CITATION_BUTTON_TEXT = r"""
@misc{open-llm-leaderboard,
  author = {Edward Beeching, ClΓ©mentine Fourrier, Nathan Habib, 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}
}"""