Spaces:
Running
Running
File size: 20,018 Bytes
e5ee086 3c69c38 e5ee086 f728b4f 0e13466 f728b4f 0e13466 f728b4f 3c69c38 e5ee086 0e13466 e5ee086 e6aadde e5ee086 3c69c38 e5ee086 f728b4f 4f4b508 f7aa050 4f4b508 9b7ca38 4f4b508 9b7ca38 f728b4f f7aa050 4f4b508 f728b4f 7853dd1 09c4409 0f6dbd0 5ec15e0 5155380 565a777 4c295e7 565a777 7853dd1 a4ad90d 7853dd1 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 |
from dataclasses import dataclass
from enum import Enum
@dataclass
class Task:
benchmark: str
metric: str
col_name: str
# Select your tasks here
# ---------------------------------------------------
class Tasks(Enum):
# task_key in the json file, metric_key in the json file, name to display in the leaderboard
task0 = Task("main", "acc", "Overall")
NUM_FEWSHOT = 0 # Change with your few shot
# ---------------------------------------------------
# Your leaderboard name
TITLE = """<h1 align="center" id="space-title">π©βπ«Invalsi Leaderboard
</h1>"""
"""
<img src="https://huggingface.co/spaces/Crisp-Unimib/INVALSIbenchmark/resolve/main/src/logo-crisp-eng-retina.png" height="800" width="150" style="vertical-align: middle;">
"""
# What does your leaderboard evaluate?
INTRODUCTION_TEXT = """
Welcome into <a href="https://crispresearch.it/"><b>CRISP Bicocca</b></a> Invalsi Leaderboard page!
We adapted the INVALSI benchmark for automated LLM evaluation, which involves rigorous adaptation of the test format to suit automated processing while retaining the essence of the original tests. In this leaderboard, we provide a detailed assessment of current LLMs, offering a crucial reference point for the academic community.
Researchers are invited to submit their models for ongoing evaluation, ensuring the benchmark remains a current and valuable resource.
For more information on the benchmark, please refer to our arXiv paper <a href="https://arxiv.org/abs/2406.17535"><b>here</b></a> and read the "About" section.
"""
# Which evaluations are you running? how can people reproduce what you have?
LLM_BENCHMARKS_TEXT = f"""
## INVALSI Benchmark
The INVALSI (National Institute for the Evaluation of the Education and Training System) test has been essential to Italy's educational assessment framework since the 2005-2006 academic year. It aims to evaluate students' competencies across various subjects, including Italian language and mathematics, at multiple stages of the education system in a standardised manner. The test's primary objective is to assess students' linguistic proficiency, a crucial skill that schools are tasked with developing as per national educational guidelines.
Since the test covers a wide range of linguistic and comprehension skills, using it to evaluate LLMs can provide a detailed view of a model's proficiency in handling real-world, nuanced language tasks designed for human learners. The test's structured and standardised nature makes it an excellent benchmark for comparing different LLMs with questions culturally and contextually relevant to Italian speakers. However, our findings are relevant across all languages since they assess various general capabilities, such as word formation and text comprehension abilities. Also, since it is designed for different educational stages, it provides varied levels of complexity and challenge. This aspect can gauge an LLM's capability at different difficulty levels, reflecting its potential scalability and adaptability across simpler to more complex linguistic tasks.
### Data Collection
We have collected from public sources 58 unique tests, divided into 141 unique units, with 2114 questions and 2808 unique items. Some questions are subdivided into multiple items, each requiring a specific answer.
The data has been gathered from the <a href="https://www.gestinv.it/Index.aspx"><b>Gestinv</b></a> database.
Questions' formatting is sometimes not adequately structured for LLM evaluation; for instance, it is sometimes impossible to automatically transcribe the questions into structured fields, necessitating further inspection of images and PDFs. For this reason, we also collected corresponding PDF files and images. Manual inspection was required to ensure accuracy. In cases where questions involved graphical elements, we modified them into an appropriate multiple-choice format. For example, if the task required a student to find and underscore a word, we reformulated the question to allow selection from multiple choices. Similarly, if the task involved drawing a line between two groups of conceptsβa common task for younger studentsβwe rephrased it to involve choosing the correct association from given options. Generally, we aimed to adapt the questions to a format that allows the model to select the correct answer from a pool of choices if it aligns with the original type of question.
### Dataset Characteristics
<ul>
<li> We have selected <b>10 tests</b> comprising <b>31 unique units</b>, <b>409 questions</b>, and <b>625 items</b> from the above data. A test comprises <b>two or more different units</b>; each question can have more than one item to answer. The sample of tests has been chosen by manual inspection, aiming to include different grades and years and avoiding those with questions that require inspecting an image or contained questions that would be hard to reformulate for language model comprehension. <br> <br>
<table style="font-size: 18px; border-collapse: collapse; width: 60%; margin: auto; border: 1px solid black;">
<caption style="caption-side: top; text-align: center; font-weight: bold;">
Distribution of tests, questions and items by educational grade.
</caption>
<thead>
<tr style="background-color: #f2f2f2;">
<th style="border: 1px solid black; padding: 8px; text-align: left;">School Grade</th>
<th style="border: 1px solid black; padding: 8px; text-align: left;"># Tests</th>
<th style="border: 1px solid black; padding: 8px; text-align: left;"># Questions</th>
<th style="border: 1px solid black; padding: 8px; text-align: left;"># Items</th>
</tr>
</thead>
<tbody>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">2nd Grade (Primary School)</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">2</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">34</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">72</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">5th Grade (Primary School)</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">2</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">75</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">117</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">6th Grade (Middle School)</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">1</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">87</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">118</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">8th Grade (Middle School)</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">2</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">86</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">88</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">10th Grade (High School)</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">2</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">77</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">139</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">13th Grade (High School)</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">1</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">50</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">91</td>
</tr>
</tbody>
</table>
</li>
<br> <br>
<li> Each question in the benchmark is labelled with the specific <b>lexical macro area</b> it aims to assess <br> <br>
<table style="font-size: 18px; border-collapse: collapse; width: 80%; margin: auto; border: 1px solid black;">
<caption style="caption-side: top; text-align: center; font-weight: bold;">
Distribution of questions by section and macro area.
</caption>
<thead>
<tr style="background-color: #f2f2f2;">
<th style="border: 1px solid black; padding: 8px; text-align: left;">Section</th>
<th style="border: 1px solid black; padding: 8px; text-align: left; width: 3.5cm;">Macro Area</th>
<th style="border: 1px solid black; padding: 8px; text-align: right;"># Questions</th>
</tr>
</thead>
<tbody>
<tr>
<td rowspan="3" style="border: 1px solid black; padding: 8px; text-align: left; width: 1.3cm;">Text comprehension</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">Reconstruct the meaning of the text, locally or globally</td>
<td style="border: 1px solid black; padding: 8px; text-align: right;">179 (43.8%)</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">Locate and identify information within the text</td>
<td style="border: 1px solid black; padding: 8px; text-align: right;">108 (26.4%)</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">Reflect on the content or form of the text, <br> locally or globally, and evaluate them</td>
<td style="border: 1px solid black; padding: 8px; text-align: right;">33 (8.1%)</td>
</tr>
<tr>
<td rowspan="6" style="border: 1px solid black; padding: 8px; text-align: left; width: 1.3cm;">Reflection on the language</td>
<td style="border: 1px solid black; padding: 8px; text-align: left;">Lexicon and semantics</td>
<td style="border: 1px solid black; padding: 8px; text-align: right;">29 (7.1%)</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">Morphology</td>
<td style="border: 1px solid black; padding: 8px; text-align: right;">24 (5.9%)</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">Syntax</td>
<td style="border: 1px solid black; padding: 8px; text-align: right;">19 (4.6%)</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">Word formation</td>
<td style="border: 1px solid black; padding: 8px; text-align: right;">7 (1.7%)</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">Textuality and pragmatics</td>
<td style="border: 1px solid black; padding: 8px; text-align: right;">6 (1.5%)</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 8px; text-align: left;">Spelling</td>
<td style="border: 1px solid black; padding: 8px; text-align: right;">4 (1.0%)</td>
</tr>
</tbody>
</table>
</li>
<br> <br>
<li> The questions come in five distinct formats <br> <br>
<table style="width: 90%; margin: auto; border-collapse: collapse; font-size: 18px; border: 1px solid black;">
<caption style="caption-side: top; text-align: center; font-weight: bold;">
Combined Legend of Question Types and Distribution by Format
</caption>
<thead>
<tr style="background-color: #f2f2f2;">
<th style="border: 1px solid black; padding: 10px; text-align: center;">Question Type</th>
<th style="border: 1px solid black; padding: 10px; text-align: center;">Description</th>
<th style="border: 1px solid black; padding: 10px; text-align: center;"># Questions</th>
<th style="border: 1px solid black; padding: 10px; text-align: center;"># Items</th>
</tr>
</thead>
<tbody>
<tr>
<td style="border: 1px solid black; padding: 10px; text-align: left;"><i>Multiple Choice (MC)</i></td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">Composed of a question with several answer options, among which only one is correct. The most common question format.</td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">334 (81.7%)</td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">340</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 10px; text-align: left;"><i>Multiple Complex Choice (MCC)</i></td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">Composed of input questions and multiple items to answer. Requires all items to be correctly answered.</td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">38 (9.3%)</td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">228</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 10px; text-align: left;"><i>Unique Response (RU)</i></td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">Open-ended questions with only one correct answer, sometimes allowing limited variants.</td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">33 (8.1%)</td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">50</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 10px; text-align: left;"><i>Short Response (RB)</i></td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">Specific sub-type of RU questions requiring a single identifying word as the answer.</td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">1 (0.2%)</td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">1</td>
</tr>
<tr>
<td style="border: 1px solid black; padding: 10px; text-align: left;"><i>Cloze (CL)</i></td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">Text with missing words to be filled in, can be open-ended or closed-ended.</td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">3 (0.7%)</td>
<td style="border: 1px solid black; padding: 10px; text-align: left;">6</td>
</tr>
</tbody>
</table>
</il>
</ul>
"""
EVALUATION_QUEUE_TEXT = """
## Evaluation Form
Please submit your model for evaluation by filling out the form located [here](https://forms.gle/4d6FYudxCVoTvj1W9).
## 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 π€
"""
CITATION_BUTTON_LABEL = "Copy the following snippet to cite these results"
CITATION_BUTTON_TEXT = r"""
@misc{mercorio2024disceautdeficereevaluating,
title={Disce aut Deficere: Evaluating LLMs Proficiency on the INVALSI Italian Benchmark},
author={Fabio Mercorio and Mario Mezzanzanica and Daniele Potertì and Antonio Serino and Andrea Seveso},
year={2024},
eprint={2406.17535},
archivePrefix={arXiv},
primaryClass={cs.CL}
url={https://arxiv.org/abs/2406.17535},
}
"""
QUESTION_FORMAT_TEXT = """
Performance (accuracy %) comparison of AI models across school grades and question formats for grades 2 to 13.
Grade 2 and 5 are part of primary school, grade 6 and 8 are part of middle school, and grade 10 and 13 are part of high school.
Due to the stratification and the limited number of questions in some categories, extreme values such as 100 or 0 are more attainable in the sections with few items.
The question formats are abbreviated as:
- _Multiple Choice (MC)_: composed of a question with several answer options, among which only one is correct.
- _Multiple Complex Choice (MCC)_: composed of input questions and multiple items to answer.
- _Unique Response (RU)_: involves open-ended questions in which there are no options or suggestions and where only one answer is considered correct (with sometimes a limited number of possible variants).
- _Short Response (RB)_: is a specific sub-type of RU questions in which the answer is short, tending to consist of a single identifying word.
- _Cloze (CL)_: can be open-ended or closed-ended. In both cases, a text with missing words is provided. The aim is to complete the text by inserting the correct missing terms.
"""
MACRO_AREA_TEXT = """"
Performance (accuracy %) comparison of AI models across semantic macro areas.
The number of questions for each category is indicated in the table headers.
Categories are abbreviated as:
* _LI_: Locate and identify information within the text.
* _RM_: Reconstruct the meaning of the text, locally or globally.
* _RC_: Reflect on the content or form of the text, locally or globally, and evaluate them.
* _WF_: Word formation.
* _LS_: Lexicon and semantics.
* _MO_: Morphology.
* _SP_: Spelling.
* _SY_: Syntax.
* _TP_: Textuality and pragmatics.
"""
CHANGELOG_TEXT = """
## Changelog
### 24/07/2024
- π New model: meta-llama/Meta-Llama-3.1-405B-Instruct
- π New model: meta-llama/Meta-Llama-3.1-70B-Instruct
- π New model: meta-llama/Meta-Llama-3.1-8B-Instruct
- π New model: google/gemma-2-27b-it
- π New model: openai/gpt-4o-mini
- π New model: cognitivecomputations/dolphin-2.9.1-llama-3-70b
- π New model: mistralai/Mistral-Nemo-Instruct-2407
- π Updated model: openai/gpt-4-turbo
- π Updated model: openai/gpt-4o
### 05/07/2024
- π New model: microsoft/phi-3-medium-128k-instruct
- π New model: microsoft/phi-3-mini-128k-instruct
### 02/07/2024
- π New model: DeepMount00/Llama-3-8b-Ita
### 30/06/2024
- π New model: google/gemma-2-9b-it
### 28/06/2024
- π New model: ludocomito/Minerva_3B_Ties_1.0
- π New model: mii-llm/maestrale-chat-v0.4-beta
- π New model: Qwen/Qwen2-72B-Instruct
- π New model: nvidia/Nemotron-4-340B-Instruct
### 27/06/2024
- π Added changelog.
- π New model: sapienzanlp/modello-italia-9b
- π New model: mixtral-8x22b-instruct
- π New model: claude-3.5-sonnet
### 26/06/2024
- π Initial leaderboard launch.
- π Added evaluation metrics for the INVALSI benchmark.
- π Added leaderboard introduction.
- β Added INVALSI benchmark description.
- π Added evaluation queue form.
- π¬ Added citation button.
- π Added question format and macro area descriptions.
"""
|