Spaces:
Running
Running
Commit ·
bf183ba
1
Parent(s): f7e0323
display ViDoRe V3 tab
Browse files
app.py
CHANGED
|
@@ -80,107 +80,21 @@ def main():
|
|
| 80 |
|
| 81 |
with gr.Blocks(css=css) as block:
|
| 82 |
with gr.Tabs():
|
| 83 |
-
with gr.TabItem("ViDoRe
|
| 84 |
-
gr.Markdown("# ViDoRe:
|
| 85 |
-
gr.
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
"""
|
| 93 |
-
)
|
| 94 |
-
datasets_columns_1 = list(data_benchmark_1.columns[4:])
|
| 95 |
-
|
| 96 |
-
with gr.Row():
|
| 97 |
-
metric_dropdown_1 = gr.Dropdown(choices=METRICS, value=initial_metric, label="Select Metric")
|
| 98 |
-
research_textbox_1 = gr.Textbox(
|
| 99 |
-
placeholder="🔍 Search Models... [press enter]",
|
| 100 |
-
label="Filter Models by Name",
|
| 101 |
-
)
|
| 102 |
-
column_checkboxes_1 = gr.CheckboxGroup(
|
| 103 |
-
choices=datasets_columns_1, value=datasets_columns_1, label="Select Columns to Display"
|
| 104 |
-
)
|
| 105 |
-
|
| 106 |
-
with gr.Row():
|
| 107 |
-
datatype_1 = ["number", "markdown"] + ["number"] * (num_datasets_1 + 1)
|
| 108 |
-
dataframe_1 = gr.Dataframe(data_benchmark_1, datatype=datatype_1, type="pandas")
|
| 109 |
-
|
| 110 |
-
def update_data_1(metric, search_term, selected_columns):
|
| 111 |
-
model_handler.get_vidore_data(metric)
|
| 112 |
-
data = model_handler.render_df(metric, benchmark_version=1)
|
| 113 |
-
data = add_rank_and_format(data, benchmark_version=1, selected_columns=selected_columns)
|
| 114 |
-
data = filter_models(data, search_term)
|
| 115 |
-
if selected_columns:
|
| 116 |
-
data = data[["Rank", "Model", "Model Size (Million Parameters)", "Average"] + selected_columns]
|
| 117 |
-
return data
|
| 118 |
-
|
| 119 |
-
with gr.Row():
|
| 120 |
-
refresh_button_1 = gr.Button("Refresh")
|
| 121 |
-
refresh_button_1.click(
|
| 122 |
-
get_refresh_function(model_handler, benchmark_version=1),
|
| 123 |
-
inputs=[metric_dropdown_1],
|
| 124 |
-
outputs=dataframe_1,
|
| 125 |
-
concurrency_limit=20,
|
| 126 |
-
)
|
| 127 |
-
|
| 128 |
-
# Automatically refresh the dataframe when the dropdown value changes
|
| 129 |
-
metric_dropdown_1.change(
|
| 130 |
-
get_refresh_function(model_handler, benchmark_version=1),
|
| 131 |
-
inputs=[metric_dropdown_1],
|
| 132 |
-
outputs=dataframe_1,
|
| 133 |
-
)
|
| 134 |
-
research_textbox_1.submit(
|
| 135 |
-
lambda metric, search_term, selected_columns: update_data_1(metric, search_term, selected_columns),
|
| 136 |
-
inputs=[metric_dropdown_1, research_textbox_1, column_checkboxes_1],
|
| 137 |
-
outputs=dataframe_1,
|
| 138 |
-
)
|
| 139 |
-
column_checkboxes_1.change(
|
| 140 |
-
lambda metric, search_term, selected_columns: update_data_1(metric, search_term, selected_columns),
|
| 141 |
-
inputs=[metric_dropdown_1, research_textbox_1, column_checkboxes_1],
|
| 142 |
-
outputs=dataframe_1,
|
| 143 |
-
)
|
| 144 |
-
|
| 145 |
-
gr.Markdown(
|
| 146 |
-
f"""
|
| 147 |
-
- **Total Datasets**: {num_datasets_1}
|
| 148 |
-
- **Total Scores**: {num_scores_1}
|
| 149 |
-
- **Total Models**: {num_models_1}
|
| 150 |
-
"""
|
| 151 |
-
+ r"""
|
| 152 |
-
Please consider citing:
|
| 153 |
-
|
| 154 |
-
```bibtex
|
| 155 |
-
@misc{faysse2024colpaliefficientdocumentretrieval,
|
| 156 |
-
title={ColPali: Efficient Document Retrieval with Vision Language Models},
|
| 157 |
-
author={Manuel Faysse and Hugues Sibille and Tony Wu and Bilel Omrani and Gautier Viaud and Céline Hudelot and Pierre Colombo},
|
| 158 |
-
year={2024},
|
| 159 |
-
eprint={2407.01449},
|
| 160 |
-
archivePrefix={arXiv},
|
| 161 |
-
primaryClass={cs.IR},
|
| 162 |
-
url={https://arxiv.org/abs/2407.01449},
|
| 163 |
-
}
|
| 164 |
-
@misc{macé2025vidorebenchmarkv2raising,
|
| 165 |
-
title={ViDoRe Benchmark V2: Raising the Bar for Visual Retrieval},
|
| 166 |
-
author={Quentin Macé and António Loison and Manuel Faysse},
|
| 167 |
-
year={2025},
|
| 168 |
-
eprint={2505.17166},
|
| 169 |
-
archivePrefix={arXiv},
|
| 170 |
-
primaryClass={cs.IR},
|
| 171 |
-
url={https://arxiv.org/abs/2505.17166},
|
| 172 |
-
}
|
| 173 |
-
|
| 174 |
-
```
|
| 175 |
-
"""
|
| 176 |
-
)
|
| 177 |
-
with gr.TabItem("ViDoRe V2"):
|
| 178 |
gr.Markdown("# ViDoRe V2: A new visual Document Retrieval Benchmark 📚🔍")
|
| 179 |
gr.Markdown("### A harder dataset benchmark for visual document retrieval 👀")
|
| 180 |
|
| 181 |
gr.Markdown(
|
| 182 |
"""
|
| 183 |
-
Visual Document Retrieval Benchmark 2 leaderboard. To submit results, refer to the corresponding tab.
|
| 184 |
|
| 185 |
Refer to the [ColPali paper](https://arxiv.org/abs/2407.01449) for details on metrics and models.
|
| 186 |
"""
|
|
@@ -223,7 +137,7 @@ def main():
|
|
| 223 |
with gr.Row():
|
| 224 |
gr.Markdown(
|
| 225 |
"""
|
| 226 |
-
**Note**: For now, all models were evaluated using the vidore-benchmark package and custom retrievers on our side.
|
| 227 |
Those numbers are not numbers obtained from the organisations that released those models.
|
| 228 |
"""
|
| 229 |
)
|
|
@@ -256,150 +170,146 @@ def main():
|
|
| 256 |
|
| 257 |
```bibtex
|
| 258 |
@misc{faysse2024colpaliefficientdocumentretrieval,
|
| 259 |
-
title={ColPali: Efficient Document Retrieval with Vision Language Models},
|
| 260 |
author={Manuel Faysse and Hugues Sibille and Tony Wu and Bilel Omrani and Gautier Viaud and Céline Hudelot and Pierre Colombo},
|
| 261 |
year={2024},
|
| 262 |
eprint={2407.01449},
|
| 263 |
archivePrefix={arXiv},
|
| 264 |
primaryClass={cs.IR},
|
| 265 |
-
url={https://arxiv.org/abs/2407.01449},
|
| 266 |
}
|
| 267 |
|
| 268 |
@misc{macé2025vidorebenchmarkv2raising,
|
| 269 |
-
title={ViDoRe Benchmark V2: Raising the Bar for Visual Retrieval},
|
| 270 |
author={Quentin Macé and António Loison and Manuel Faysse},
|
| 271 |
year={2025},
|
| 272 |
eprint={2505.17166},
|
| 273 |
archivePrefix={arXiv},
|
| 274 |
primaryClass={cs.IR},
|
| 275 |
-
url={https://arxiv.org/abs/2505.17166},
|
| 276 |
}
|
| 277 |
```
|
| 278 |
"""
|
| 279 |
)
|
| 280 |
-
with gr.TabItem("📚 Submit your model"):
|
| 281 |
-
gr.Markdown("# How to Submit a New Model to the Leaderboard")
|
| 282 |
-
gr.Markdown(
|
| 283 |
-
"""
|
| 284 |
-
To submit a new model to the ViDoRe leaderboard, follow these steps:
|
| 285 |
-
|
| 286 |
-
1. **Evaluate your model**:
|
| 287 |
-
- Follow the evaluation procedure provided in the [ViDoRe GitHub repository](https://github.com/illuin-tech/vidore-benchmark/) that uses MTEB.
|
| 288 |
-
|
| 289 |
-
2. **Format your submission file**:
|
| 290 |
-
- Add the generated files to [MTEB results](https://github.com/embeddings-benchmark/results) project. Check the [Colpali results](https://github.com/embeddings-benchmark/results/tree/main/results/vidore__colpali-v1.3/1b5c8929330df1a66de441a9b5409a878f0de5b0) for an example.
|
| 291 |
-
|
| 292 |
-
And you're done! Your model will appear on the leaderboard when you click refresh! Once the space
|
| 293 |
-
gets rebooted, it will appear on startup.
|
| 294 |
|
| 295 |
-
|
| 296 |
-
|
| 297 |
-
"""
|
| 298 |
-
)
|
| 299 |
-
with gr.TabItem("Deprecated ViDoRe V1"):
|
| 300 |
-
gr.Markdown(
|
| 301 |
-
"## <span style='color:red'>Deprecation notice: This leaderboard contains the results computed with the "
|
| 302 |
-
"[vidore-benchmark](https://github.com/illuin-tech/vidore-benchmark) package, "
|
| 303 |
-
"which is no longer maintained. Results should be computed using the "
|
| 304 |
-
"[mteb](https://github.com/embeddings-benchmark/mteb) package as described "
|
| 305 |
-
"[here](https://github.com/illuin-tech/vidore-benchmark/blob/main/README.md).</span>"
|
| 306 |
-
)
|
| 307 |
-
gr.Markdown("## <span style='color:red'>Missing results in the new leaderboard are being added as they are re-computed.</span>")
|
| 308 |
-
gr.Markdown("# <span style='color:red'>[Deprecated]</span> ViDoRe: The Visual Document Retrieval Benchmark 1 📚🔍")
|
| 309 |
gr.Markdown("### From the paper - ColPali: Efficient Document Retrieval with Vision Language Models 👀")
|
| 310 |
|
| 311 |
gr.Markdown(
|
| 312 |
"""
|
| 313 |
-
Visual Document Retrieval Benchmark 1 leaderboard. To submit results, refer to the corresponding tab.
|
| 314 |
|
| 315 |
Refer to the [ColPali paper](https://arxiv.org/abs/2407.01449) for details on metrics, tasks and models.
|
| 316 |
"""
|
| 317 |
)
|
| 318 |
-
|
| 319 |
|
| 320 |
with gr.Row():
|
| 321 |
-
|
| 322 |
-
|
| 323 |
placeholder="🔍 Search Models... [press enter]",
|
| 324 |
label="Filter Models by Name",
|
| 325 |
)
|
| 326 |
-
|
| 327 |
-
choices=
|
| 328 |
)
|
| 329 |
|
| 330 |
with gr.Row():
|
| 331 |
-
|
| 332 |
-
|
| 333 |
|
| 334 |
-
def
|
| 335 |
-
|
| 336 |
-
data =
|
| 337 |
data = add_rank_and_format(data, benchmark_version=1, selected_columns=selected_columns)
|
| 338 |
data = filter_models(data, search_term)
|
| 339 |
-
# data = remove_duplicates(data) # Add this line
|
| 340 |
if selected_columns:
|
| 341 |
-
data = data[["Rank", "Model", "Average"] + selected_columns]
|
| 342 |
return data
|
| 343 |
|
| 344 |
with gr.Row():
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
inputs=[
|
| 349 |
-
outputs=
|
| 350 |
concurrency_limit=20,
|
| 351 |
)
|
| 352 |
|
| 353 |
# Automatically refresh the dataframe when the dropdown value changes
|
| 354 |
-
|
| 355 |
-
|
| 356 |
-
inputs=[
|
| 357 |
-
outputs=
|
| 358 |
)
|
| 359 |
-
|
| 360 |
-
lambda metric, search_term, selected_columns:
|
| 361 |
-
inputs=[
|
| 362 |
-
outputs=
|
| 363 |
)
|
| 364 |
-
|
| 365 |
-
lambda metric, search_term, selected_columns:
|
| 366 |
-
inputs=[
|
| 367 |
-
outputs=
|
| 368 |
)
|
| 369 |
|
| 370 |
gr.Markdown(
|
| 371 |
f"""
|
| 372 |
-
- **Total Datasets**: {
|
| 373 |
-
- **Total Scores**: {
|
| 374 |
-
- **Total Models**: {
|
| 375 |
"""
|
| 376 |
+ r"""
|
| 377 |
Please consider citing:
|
| 378 |
|
| 379 |
```bibtex
|
| 380 |
@misc{faysse2024colpaliefficientdocumentretrieval,
|
| 381 |
-
title={ColPali: Efficient Document Retrieval with Vision Language Models},
|
| 382 |
author={Manuel Faysse and Hugues Sibille and Tony Wu and Bilel Omrani and Gautier Viaud and Céline Hudelot and Pierre Colombo},
|
| 383 |
year={2024},
|
| 384 |
eprint={2407.01449},
|
| 385 |
archivePrefix={arXiv},
|
| 386 |
primaryClass={cs.IR},
|
| 387 |
-
url={https://arxiv.org/abs/2407.01449},
|
| 388 |
}
|
| 389 |
-
|
| 390 |
@misc{macé2025vidorebenchmarkv2raising,
|
| 391 |
-
title={ViDoRe Benchmark V2: Raising the Bar for Visual Retrieval},
|
| 392 |
author={Quentin Macé and António Loison and Manuel Faysse},
|
| 393 |
year={2025},
|
| 394 |
eprint={2505.17166},
|
| 395 |
archivePrefix={arXiv},
|
| 396 |
primaryClass={cs.IR},
|
| 397 |
-
url={https://arxiv.org/abs/2505.17166},
|
| 398 |
}
|
|
|
|
| 399 |
```
|
| 400 |
"""
|
| 401 |
)
|
| 402 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 403 |
gr.Markdown(
|
| 404 |
"## <span style='color:red'>Deprecation notice: This leaderboard contains the results computed with the "
|
| 405 |
"[vidore-benchmark](https://github.com/illuin-tech/vidore-benchmark) package, "
|
|
@@ -413,7 +323,7 @@ def main():
|
|
| 413 |
|
| 414 |
gr.Markdown(
|
| 415 |
"""
|
| 416 |
-
Visual Document Retrieval Benchmark 2 leaderboard. To submit results, refer to the corresponding tab.
|
| 417 |
|
| 418 |
Refer to the [ColPali paper](https://arxiv.org/abs/2407.01449) for details on metrics and models.
|
| 419 |
"""
|
|
@@ -456,7 +366,7 @@ def main():
|
|
| 456 |
with gr.Row():
|
| 457 |
gr.Markdown(
|
| 458 |
"""
|
| 459 |
-
**Note**: For now, all models were evaluated using the vidore-benchmark package and custom retrievers on our side.
|
| 460 |
Those numbers are not numbers obtained from the organisations that released those models.
|
| 461 |
"""
|
| 462 |
)
|
|
@@ -489,30 +399,133 @@ def main():
|
|
| 489 |
|
| 490 |
```bibtex
|
| 491 |
@misc{faysse2024colpaliefficientdocumentretrieval,
|
| 492 |
-
title={ColPali: Efficient Document Retrieval with Vision Language Models},
|
| 493 |
author={Manuel Faysse and Hugues Sibille and Tony Wu and Bilel Omrani and Gautier Viaud and Céline Hudelot and Pierre Colombo},
|
| 494 |
year={2024},
|
| 495 |
eprint={2407.01449},
|
| 496 |
archivePrefix={arXiv},
|
| 497 |
primaryClass={cs.IR},
|
| 498 |
-
url={https://arxiv.org/abs/2407.01449},
|
| 499 |
}
|
| 500 |
|
| 501 |
@misc{macé2025vidorebenchmarkv2raising,
|
| 502 |
-
title={ViDoRe Benchmark V2: Raising the Bar for Visual Retrieval},
|
| 503 |
author={Quentin Macé and António Loison and Manuel Faysse},
|
| 504 |
year={2025},
|
| 505 |
eprint={2505.17166},
|
| 506 |
archivePrefix={arXiv},
|
| 507 |
primaryClass={cs.IR},
|
| 508 |
-
url={https://arxiv.org/abs/2505.17166},
|
| 509 |
}
|
| 510 |
```
|
| 511 |
"""
|
| 512 |
)
|
| 513 |
|
| 514 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 515 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 516 |
|
| 517 |
if __name__ == "__main__":
|
| 518 |
main()
|
|
|
|
| 80 |
|
| 81 |
with gr.Blocks(css=css) as block:
|
| 82 |
with gr.Tabs():
|
| 83 |
+
with gr.TabItem("3️⃣ ViDoRe V3"):
|
| 84 |
+
gr.Markdown("# ViDoRe V3: a comprehensive evaluation of retrieval for enterprise use-cases 📚🔍")
|
| 85 |
+
with gr.Row(variant="panel"):
|
| 86 |
+
gr.Markdown("""
|
| 87 |
+
### ⚠️ To access the ViDoRe V3 results, please refer directly to the [MTEB Leaderboard](https://huggingface.co/spaces/mteb/leaderboard).
|
| 88 |
+
**ViDoRe V3 is fully integrated into MTEB, which provides a unified platform for evaluating embedding models across various tasks, including document retrieval.**
|
| 89 |
+
**We decided to display ViDoRe V3 results directly on MTEB to leverage its extensive features and community.**
|
| 90 |
+
""")
|
| 91 |
+
with gr.TabItem("2️⃣ ViDoRe V2"):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 92 |
gr.Markdown("# ViDoRe V2: A new visual Document Retrieval Benchmark 📚🔍")
|
| 93 |
gr.Markdown("### A harder dataset benchmark for visual document retrieval 👀")
|
| 94 |
|
| 95 |
gr.Markdown(
|
| 96 |
"""
|
| 97 |
+
Visual Document Retrieval Benchmark 2 leaderboard. To submit results, refer to the corresponding tab.
|
| 98 |
|
| 99 |
Refer to the [ColPali paper](https://arxiv.org/abs/2407.01449) for details on metrics and models.
|
| 100 |
"""
|
|
|
|
| 137 |
with gr.Row():
|
| 138 |
gr.Markdown(
|
| 139 |
"""
|
| 140 |
+
**Note**: For now, all models were evaluated using the vidore-benchmark package and custom retrievers on our side.
|
| 141 |
Those numbers are not numbers obtained from the organisations that released those models.
|
| 142 |
"""
|
| 143 |
)
|
|
|
|
| 170 |
|
| 171 |
```bibtex
|
| 172 |
@misc{faysse2024colpaliefficientdocumentretrieval,
|
| 173 |
+
title={ColPali: Efficient Document Retrieval with Vision Language Models},
|
| 174 |
author={Manuel Faysse and Hugues Sibille and Tony Wu and Bilel Omrani and Gautier Viaud and Céline Hudelot and Pierre Colombo},
|
| 175 |
year={2024},
|
| 176 |
eprint={2407.01449},
|
| 177 |
archivePrefix={arXiv},
|
| 178 |
primaryClass={cs.IR},
|
| 179 |
+
url={https://arxiv.org/abs/2407.01449},
|
| 180 |
}
|
| 181 |
|
| 182 |
@misc{macé2025vidorebenchmarkv2raising,
|
| 183 |
+
title={ViDoRe Benchmark V2: Raising the Bar for Visual Retrieval},
|
| 184 |
author={Quentin Macé and António Loison and Manuel Faysse},
|
| 185 |
year={2025},
|
| 186 |
eprint={2505.17166},
|
| 187 |
archivePrefix={arXiv},
|
| 188 |
primaryClass={cs.IR},
|
| 189 |
+
url={https://arxiv.org/abs/2505.17166},
|
| 190 |
}
|
| 191 |
```
|
| 192 |
"""
|
| 193 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 194 |
|
| 195 |
+
with gr.TabItem("1️⃣ ViDoRe V1"):
|
| 196 |
+
gr.Markdown("# ViDoRe: The Visual Document Retrieval Benchmark 1 📚🔍")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 197 |
gr.Markdown("### From the paper - ColPali: Efficient Document Retrieval with Vision Language Models 👀")
|
| 198 |
|
| 199 |
gr.Markdown(
|
| 200 |
"""
|
| 201 |
+
Visual Document Retrieval Benchmark 1 leaderboard. To submit results, refer to the corresponding tab.
|
| 202 |
|
| 203 |
Refer to the [ColPali paper](https://arxiv.org/abs/2407.01449) for details on metrics, tasks and models.
|
| 204 |
"""
|
| 205 |
)
|
| 206 |
+
datasets_columns_1 = list(data_benchmark_1.columns[4:])
|
| 207 |
|
| 208 |
with gr.Row():
|
| 209 |
+
metric_dropdown_1 = gr.Dropdown(choices=METRICS, value=initial_metric, label="Select Metric")
|
| 210 |
+
research_textbox_1 = gr.Textbox(
|
| 211 |
placeholder="🔍 Search Models... [press enter]",
|
| 212 |
label="Filter Models by Name",
|
| 213 |
)
|
| 214 |
+
column_checkboxes_1 = gr.CheckboxGroup(
|
| 215 |
+
choices=datasets_columns_1, value=datasets_columns_1, label="Select Columns to Display"
|
| 216 |
)
|
| 217 |
|
| 218 |
with gr.Row():
|
| 219 |
+
datatype_1 = ["number", "markdown"] + ["number"] * (num_datasets_1 + 1)
|
| 220 |
+
dataframe_1 = gr.Dataframe(data_benchmark_1, datatype=datatype_1, type="pandas")
|
| 221 |
|
| 222 |
+
def update_data_1(metric, search_term, selected_columns):
|
| 223 |
+
model_handler.get_vidore_data(metric)
|
| 224 |
+
data = model_handler.render_df(metric, benchmark_version=1)
|
| 225 |
data = add_rank_and_format(data, benchmark_version=1, selected_columns=selected_columns)
|
| 226 |
data = filter_models(data, search_term)
|
|
|
|
| 227 |
if selected_columns:
|
| 228 |
+
data = data[["Rank", "Model", "Model Size (Million Parameters)", "Average"] + selected_columns]
|
| 229 |
return data
|
| 230 |
|
| 231 |
with gr.Row():
|
| 232 |
+
refresh_button_1 = gr.Button("Refresh")
|
| 233 |
+
refresh_button_1.click(
|
| 234 |
+
get_refresh_function(model_handler, benchmark_version=1),
|
| 235 |
+
inputs=[metric_dropdown_1],
|
| 236 |
+
outputs=dataframe_1,
|
| 237 |
concurrency_limit=20,
|
| 238 |
)
|
| 239 |
|
| 240 |
# Automatically refresh the dataframe when the dropdown value changes
|
| 241 |
+
metric_dropdown_1.change(
|
| 242 |
+
get_refresh_function(model_handler, benchmark_version=1),
|
| 243 |
+
inputs=[metric_dropdown_1],
|
| 244 |
+
outputs=dataframe_1,
|
| 245 |
)
|
| 246 |
+
research_textbox_1.submit(
|
| 247 |
+
lambda metric, search_term, selected_columns: update_data_1(metric, search_term, selected_columns),
|
| 248 |
+
inputs=[metric_dropdown_1, research_textbox_1, column_checkboxes_1],
|
| 249 |
+
outputs=dataframe_1,
|
| 250 |
)
|
| 251 |
+
column_checkboxes_1.change(
|
| 252 |
+
lambda metric, search_term, selected_columns: update_data_1(metric, search_term, selected_columns),
|
| 253 |
+
inputs=[metric_dropdown_1, research_textbox_1, column_checkboxes_1],
|
| 254 |
+
outputs=dataframe_1,
|
| 255 |
)
|
| 256 |
|
| 257 |
gr.Markdown(
|
| 258 |
f"""
|
| 259 |
+
- **Total Datasets**: {num_datasets_1}
|
| 260 |
+
- **Total Scores**: {num_scores_1}
|
| 261 |
+
- **Total Models**: {num_models_1}
|
| 262 |
"""
|
| 263 |
+ r"""
|
| 264 |
Please consider citing:
|
| 265 |
|
| 266 |
```bibtex
|
| 267 |
@misc{faysse2024colpaliefficientdocumentretrieval,
|
| 268 |
+
title={ColPali: Efficient Document Retrieval with Vision Language Models},
|
| 269 |
author={Manuel Faysse and Hugues Sibille and Tony Wu and Bilel Omrani and Gautier Viaud and Céline Hudelot and Pierre Colombo},
|
| 270 |
year={2024},
|
| 271 |
eprint={2407.01449},
|
| 272 |
archivePrefix={arXiv},
|
| 273 |
primaryClass={cs.IR},
|
| 274 |
+
url={https://arxiv.org/abs/2407.01449},
|
| 275 |
}
|
|
|
|
| 276 |
@misc{macé2025vidorebenchmarkv2raising,
|
| 277 |
+
title={ViDoRe Benchmark V2: Raising the Bar for Visual Retrieval},
|
| 278 |
author={Quentin Macé and António Loison and Manuel Faysse},
|
| 279 |
year={2025},
|
| 280 |
eprint={2505.17166},
|
| 281 |
archivePrefix={arXiv},
|
| 282 |
primaryClass={cs.IR},
|
| 283 |
+
url={https://arxiv.org/abs/2505.17166},
|
| 284 |
}
|
| 285 |
+
|
| 286 |
```
|
| 287 |
"""
|
| 288 |
)
|
| 289 |
+
|
| 290 |
+
with gr.TabItem("📚 Submit your model"):
|
| 291 |
+
gr.Markdown("# How to Submit a New Model to the Leaderboard")
|
| 292 |
+
gr.Markdown(
|
| 293 |
+
"""
|
| 294 |
+
To submit a new model to the ViDoRe leaderboard, follow these steps:
|
| 295 |
+
|
| 296 |
+
1. **Evaluate your model**:
|
| 297 |
+
- Follow the evaluation procedure provided in the [ViDoRe GitHub repository](https://github.com/illuin-tech/vidore-benchmark/) that uses MTEB.
|
| 298 |
+
|
| 299 |
+
2. **Format your submission file**:
|
| 300 |
+
- Add the generated files to [MTEB results](https://github.com/embeddings-benchmark/results) project. Check the [Colpali results](https://github.com/embeddings-benchmark/results/tree/main/results/vidore__colpali-v1.3/1b5c8929330df1a66de441a9b5409a878f0de5b0) for an example.
|
| 301 |
+
|
| 302 |
+
And you're done! Your model will appear on the leaderboard when you click refresh! Once the space
|
| 303 |
+
gets rebooted, it will appear on startup.
|
| 304 |
+
|
| 305 |
+
Note: For proper hyperlink redirection, please ensure that your model repository name is in
|
| 306 |
+
kebab-case, e.g. `my-model-name`.
|
| 307 |
+
"""
|
| 308 |
+
)
|
| 309 |
+
|
| 310 |
+
### Deprecated Tabs ###
|
| 311 |
+
|
| 312 |
+
with gr.TabItem("⚠️ Deprecated ViDoRe V2"):
|
| 313 |
gr.Markdown(
|
| 314 |
"## <span style='color:red'>Deprecation notice: This leaderboard contains the results computed with the "
|
| 315 |
"[vidore-benchmark](https://github.com/illuin-tech/vidore-benchmark) package, "
|
|
|
|
| 323 |
|
| 324 |
gr.Markdown(
|
| 325 |
"""
|
| 326 |
+
Visual Document Retrieval Benchmark 2 leaderboard. To submit results, refer to the corresponding tab.
|
| 327 |
|
| 328 |
Refer to the [ColPali paper](https://arxiv.org/abs/2407.01449) for details on metrics and models.
|
| 329 |
"""
|
|
|
|
| 366 |
with gr.Row():
|
| 367 |
gr.Markdown(
|
| 368 |
"""
|
| 369 |
+
**Note**: For now, all models were evaluated using the vidore-benchmark package and custom retrievers on our side.
|
| 370 |
Those numbers are not numbers obtained from the organisations that released those models.
|
| 371 |
"""
|
| 372 |
)
|
|
|
|
| 399 |
|
| 400 |
```bibtex
|
| 401 |
@misc{faysse2024colpaliefficientdocumentretrieval,
|
| 402 |
+
title={ColPali: Efficient Document Retrieval with Vision Language Models},
|
| 403 |
author={Manuel Faysse and Hugues Sibille and Tony Wu and Bilel Omrani and Gautier Viaud and Céline Hudelot and Pierre Colombo},
|
| 404 |
year={2024},
|
| 405 |
eprint={2407.01449},
|
| 406 |
archivePrefix={arXiv},
|
| 407 |
primaryClass={cs.IR},
|
| 408 |
+
url={https://arxiv.org/abs/2407.01449},
|
| 409 |
}
|
| 410 |
|
| 411 |
@misc{macé2025vidorebenchmarkv2raising,
|
| 412 |
+
title={ViDoRe Benchmark V2: Raising the Bar for Visual Retrieval},
|
| 413 |
author={Quentin Macé and António Loison and Manuel Faysse},
|
| 414 |
year={2025},
|
| 415 |
eprint={2505.17166},
|
| 416 |
archivePrefix={arXiv},
|
| 417 |
primaryClass={cs.IR},
|
| 418 |
+
url={https://arxiv.org/abs/2505.17166},
|
| 419 |
}
|
| 420 |
```
|
| 421 |
"""
|
| 422 |
)
|
| 423 |
|
| 424 |
+
with gr.TabItem("⚠️ Deprecated ViDoRe V1"):
|
| 425 |
+
gr.Markdown(
|
| 426 |
+
"## <span style='color:red'>Deprecation notice: This leaderboard contains the results computed with the "
|
| 427 |
+
"[vidore-benchmark](https://github.com/illuin-tech/vidore-benchmark) package, "
|
| 428 |
+
"which is no longer maintained. Results should be computed using the "
|
| 429 |
+
"[mteb](https://github.com/embeddings-benchmark/mteb) package as described "
|
| 430 |
+
"[here](https://github.com/illuin-tech/vidore-benchmark/blob/main/README.md).</span>"
|
| 431 |
+
)
|
| 432 |
+
gr.Markdown("## <span style='color:red'>Missing results in the new leaderboard are being added as they are re-computed.</span>")
|
| 433 |
+
gr.Markdown("# <span style='color:red'>[Deprecated]</span> ViDoRe: The Visual Document Retrieval Benchmark 1 📚🔍")
|
| 434 |
+
gr.Markdown("### From the paper - ColPali: Efficient Document Retrieval with Vision Language Models 👀")
|
| 435 |
+
|
| 436 |
+
gr.Markdown(
|
| 437 |
+
"""
|
| 438 |
+
Visual Document Retrieval Benchmark 1 leaderboard. To submit results, refer to the corresponding tab.
|
| 439 |
+
|
| 440 |
+
Refer to the [ColPali paper](https://arxiv.org/abs/2407.01449) for details on metrics, tasks and models.
|
| 441 |
+
"""
|
| 442 |
+
)
|
| 443 |
+
deprecated_datasets_columns_1 = list(deprecated_data_benchmark_1.columns[3:])
|
| 444 |
+
|
| 445 |
+
with gr.Row():
|
| 446 |
+
deprecated_metric_dropdown_1 = gr.Dropdown(choices=METRICS, value=initial_metric, label="Select Metric")
|
| 447 |
+
deprecated_research_textbox_1 = gr.Textbox(
|
| 448 |
+
placeholder="🔍 Search Models... [press enter]",
|
| 449 |
+
label="Filter Models by Name",
|
| 450 |
+
)
|
| 451 |
+
deprecated_column_checkboxes_1 = gr.CheckboxGroup(
|
| 452 |
+
choices=deprecated_datasets_columns_1, value=deprecated_datasets_columns_1, label="Select Columns to Display"
|
| 453 |
+
)
|
| 454 |
+
|
| 455 |
+
with gr.Row():
|
| 456 |
+
deprecated_datatype_1 = ["number", "markdown"] + ["number"] * (deprecated_num_datasets_1 + 1)
|
| 457 |
+
deprecated_dataframe_1 = gr.Dataframe(deprecated_data_benchmark_1, datatype=deprecated_datatype_1, type="pandas")
|
| 458 |
+
|
| 459 |
+
def deprecated_update_data_1(metric, search_term, selected_columns):
|
| 460 |
+
deprecated_model_handler.get_vidore_data(metric)
|
| 461 |
+
data = deprecated_model_handler.render_df(metric, benchmark_version=1)
|
| 462 |
+
data = add_rank_and_format(data, benchmark_version=1, selected_columns=selected_columns)
|
| 463 |
+
data = filter_models(data, search_term)
|
| 464 |
+
# data = remove_duplicates(data) # Add this line
|
| 465 |
+
if selected_columns:
|
| 466 |
+
data = data[["Rank", "Model", "Average"] + selected_columns]
|
| 467 |
+
return data
|
| 468 |
|
| 469 |
+
with gr.Row():
|
| 470 |
+
deprecated_refresh_button_1 = gr.Button("Refresh")
|
| 471 |
+
deprecated_refresh_button_1.click(
|
| 472 |
+
deprecated_get_refresh_function(deprecated_model_handler, benchmark_version=1),
|
| 473 |
+
inputs=[deprecated_metric_dropdown_1],
|
| 474 |
+
outputs=deprecated_dataframe_1,
|
| 475 |
+
concurrency_limit=20,
|
| 476 |
+
)
|
| 477 |
+
|
| 478 |
+
# Automatically refresh the dataframe when the dropdown value changes
|
| 479 |
+
deprecated_metric_dropdown_1.change(
|
| 480 |
+
deprecated_get_refresh_function(deprecated_model_handler, benchmark_version=1),
|
| 481 |
+
inputs=[deprecated_metric_dropdown_1],
|
| 482 |
+
outputs=deprecated_dataframe_1,
|
| 483 |
+
)
|
| 484 |
+
deprecated_research_textbox_1.submit(
|
| 485 |
+
lambda metric, search_term, selected_columns: deprecated_update_data_1(metric, search_term, selected_columns),
|
| 486 |
+
inputs=[deprecated_metric_dropdown_1, deprecated_research_textbox_1, deprecated_column_checkboxes_1],
|
| 487 |
+
outputs=deprecated_dataframe_1,
|
| 488 |
+
)
|
| 489 |
+
deprecated_column_checkboxes_1.change(
|
| 490 |
+
lambda metric, search_term, selected_columns: deprecated_update_data_1(metric, search_term, selected_columns),
|
| 491 |
+
inputs=[deprecated_metric_dropdown_1, deprecated_research_textbox_1, deprecated_column_checkboxes_1],
|
| 492 |
+
outputs=deprecated_dataframe_1,
|
| 493 |
+
)
|
| 494 |
+
|
| 495 |
+
gr.Markdown(
|
| 496 |
+
f"""
|
| 497 |
+
- **Total Datasets**: {deprecated_num_datasets_1}
|
| 498 |
+
- **Total Scores**: {deprecated_num_scores_1}
|
| 499 |
+
- **Total Models**: {deprecated_num_models_1}
|
| 500 |
+
"""
|
| 501 |
+
+ r"""
|
| 502 |
+
Please consider citing:
|
| 503 |
+
|
| 504 |
+
```bibtex
|
| 505 |
+
@misc{faysse2024colpaliefficientdocumentretrieval,
|
| 506 |
+
title={ColPali: Efficient Document Retrieval with Vision Language Models},
|
| 507 |
+
author={Manuel Faysse and Hugues Sibille and Tony Wu and Bilel Omrani and Gautier Viaud and Céline Hudelot and Pierre Colombo},
|
| 508 |
+
year={2024},
|
| 509 |
+
eprint={2407.01449},
|
| 510 |
+
archivePrefix={arXiv},
|
| 511 |
+
primaryClass={cs.IR},
|
| 512 |
+
url={https://arxiv.org/abs/2407.01449},
|
| 513 |
+
}
|
| 514 |
+
|
| 515 |
+
@misc{macé2025vidorebenchmarkv2raising,
|
| 516 |
+
title={ViDoRe Benchmark V2: Raising the Bar for Visual Retrieval},
|
| 517 |
+
author={Quentin Macé and António Loison and Manuel Faysse},
|
| 518 |
+
year={2025},
|
| 519 |
+
eprint={2505.17166},
|
| 520 |
+
archivePrefix={arXiv},
|
| 521 |
+
primaryClass={cs.IR},
|
| 522 |
+
url={https://arxiv.org/abs/2505.17166},
|
| 523 |
+
}
|
| 524 |
+
```
|
| 525 |
+
"""
|
| 526 |
+
)
|
| 527 |
+
|
| 528 |
+
block.queue(max_size=10).launch(debug=True)
|
| 529 |
|
| 530 |
if __name__ == "__main__":
|
| 531 |
main()
|