yjernite HF Staff commited on
Commit
fdaa39b
Β·
verified Β·
1 Parent(s): 3f16b59

Upload app.py

Browse files
Files changed (1) hide show
  1. app.py +223 -201
app.py CHANGED
@@ -85,234 +85,256 @@ def create_app():
85
  """
86
 
87
  with gr.Blocks(css=custom_css, title="SmolLM3-3B EU Data Transparency") as app:
88
- # Banner section with images
89
- with gr.Row():
90
- with gr.Column(scale=2):
91
- try:
92
- gr.Image("eu_flag.png", height=180, show_label=False, show_download_button=False, interactive=False, container=False)
93
- except:
94
- gr.HTML('<div style="height: 120px;"></div>') # Placeholder if image not found
95
-
96
- with gr.Column(scale=1.5):
97
- gr.HTML("""
98
- <div style="text-align: center; padding: 20px;">
99
- <h1 style="color: #2c3e50; margin: 0; font-size: 3em !important;">SmolLM3-3B</h1>
100
- <h2 style="color: #667eea; margin: 10px 0 0 0; font-size: 1.5em !important;">Public Summary of Training Content</h2>
101
- </div>
102
- """)
103
-
104
- with gr.Column(scale=2):
105
- try:
106
- gr.Image("banner.png", height=180, show_label=False, show_download_button=False, interactive=False, container=False)
107
- except:
108
- gr.HTML('<div style="height: 120px;"></div>') # Placeholder if image not found
109
-
110
- gr.HTML("""
111
- <div style="text-align: center; margin-top: 40px; padding: 20px; background: #f8f9fa; border-radius: 10px;">
112
- <p style="color: #6c757d; margin: 0;">
113
- This Space contains the transparency report for the <a href="https://huggingface.co/HuggingFaceTB/SmolLM3-3B">SmolLM3-3B</a> GPAI model developped by <a href="https://huggingface.co/">Hugging Face</a> following the guidelines provided by the AI Office.<br/>
114
- For more information, see the <a href="https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models" class="dataset-link">Explanatory Notice and Template</a>
115
- </p>
116
- </div>
117
- """)
118
 
119
  with gr.Column(elem_classes=["main-container"]):
120
- # Section 1: General Information
121
- gr.HTML('<div class="section-header">1. General information</div>')
122
-
123
  with gr.Row():
124
- with gr.Column():
 
 
 
 
 
 
125
  gr.HTML("""
126
- <div class="info-box">
127
- <div class="subsection-header">1.1. Provider identification</div>
128
- <ul>
129
- <li><strong>Provider name and contact details:</strong>
130
- <ul>
131
- <li><strong>Hugging Face</strong></li>
132
- <li><strong>Website: <a href="https://huggingface.co" class="dataset-link">https://huggingface.co</a></strong></li>
133
- </ul>
134
- </li>
135
- </ul>
136
  </div>
137
  """)
 
 
 
 
 
 
138
 
139
- with gr.Column():
140
- gr.HTML("""
141
- <div class="info-box">
142
- <div class="subsection-header">1.2. Model identification</div>
143
- <ul>
144
- <li><strong>Versioned model name(s):</strong>
145
- <ul><li><strong>SmolLM3-3B</strong></li></ul>
146
- </li>
147
- <li><strong>Model dependencies:</strong>
148
- <ul><li><strong>None</strong></li></ul>
149
- </li>
150
- </ul>
151
- </div>
152
- """)
153
 
 
 
154
  gr.HTML("""
155
- <div class="info-box">
156
- <div class="subsection-header">1.3. Modalities, overall training data size and other characteristics</div>
157
- <ul>
158
- <li><strong>TEXT</strong>
159
- <ul>
160
- <li><strong>Size:</strong> <strong>more than 10 trillion tokens</strong></li>
161
- <li>The training corpus for SmolLM3 is made up of several publicly accessible large datasets containing web documents, scientific articles, software code, and synthetically generated textbooks and mathematical data for pre-training in addition to several mid-training and fine-tuning datasets to enable chat interactions, instruction-following and task-solving behaviors.</li>
162
- </ul>
163
- </li>
164
- <li><strong>Latest date of data acquisition/collection for model training:</strong>
165
- <ul>
166
- <li>The training dataset is made up of different subsets with different publication and cutoff dates. For pre-training, the earliest dataset was last updated on 4/3/2024 (Stack v2), and the latest on 2/19/2025 (FineWeb2-HQ)</li>
167
- </ul>
168
- </li>
169
- <li><strong>Description of the linguistic characteristics of the overall training data:</strong>
170
- <ul>
171
- <li>The overall training process focuses on 6 languages that are all Union languages: English, French, Spanish, German, Italian, and Portuguese. In addition, pre-training intentionally included smaller quantities of data in Mandarin Chinese, Russian, Persian, Japanese, Korean, Vietnamese, Hindi, Thai, and Greek. Other languages may have been included due to the limitations of automatic language identification in filtering stages.</li>
172
- </ul>
173
- </li>
174
- <li><strong>Other relevant characteristics of the overall training data:</strong>
175
- <ul>
176
- <li>The training data also includes software code in the programming languages included in the Stack v2 dataset (16 languages including C, Python, Java, Markdown, HTML, Shell, etc.).</li>
177
- </ul>
178
- </li>
179
- </ul>
180
  </div>
181
  """)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
182
 
183
  # Section 2: Data Sources
184
  gr.HTML('<div class="section-header">2. List of data sources</div>')
185
-
186
  gr.HTML("""
187
- <div class="info-box">
188
- <div class="subsection-header">2.1. Publicly available datasets</div>
189
- <ul>
190
- <li><strong>Have you used publicly available datasets to train the model?</strong>
191
- <ul><li><strong><span class="checkbox-yes">β˜‘ Yes</span></strong></li></ul>
192
- </li>
193
- <li><strong>If yes, specify the modality(ies) of the content covered by the datasets concerned:</strong>
194
- <ul><li><strong><span class="checkbox-yes">β˜‘ Text</span></strong></li></ul>
195
- </li>
196
- <li><strong>List of large publicly available datasets:</strong>
197
- <ul>
198
- <li>DCLM: <a href="https://hf.co/datasets/mlfoundations/dclm-baseline-1.0" class="dataset-link">https://hf.co/datasets/mlfoundations/dclm-baseline-1.0</a></li>
199
- <li>FineWeb-Edu: <a href="https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu" class="dataset-link">https://hf.co/datasets/HuggingFaceFW/fineweb-edu</a></li>
200
- <li>FineWeb2: <a href="https://huggingface.co/datasets/epfml/FineWeb2-HQ" class="dataset-link">https://hf.co/datasets/epfml/FineWeb2-HQ</a></li>
201
- <li>Stack V2: <a href="https://hf.co/datasets/bigcode/the-stack-v2" class="dataset-link">https://hf.co/datasets/bigcode/the-stack-v2</a></li>
202
- <li>pes2o: <a href="https://hf.co/datasets/allenai/peS2o" class="dataset-link">https://hf.co/datasets/allenai/peS2o</a></li>
203
- <li>SmolTalk2: <a href="https://huggingface.co/datasets/HuggingFaceTB/smoltalk2" class="dataset-link">https://hf.co/datasets/HuggingFaceTB/smoltalk2</a></li>
204
- </ul>
205
- </li>
206
- <li><strong>General description of other publicly available datasets not listed above:</strong>
207
- <ul>
208
- <li>In addition to the large datasets cited above, many additional publicly available datasets were added to target specific domains, including several math datasets made up of both web-filtered and synthetic data, Wikipedia data, "reasoning data" generated by selected large models on diverse problems, Jupyter notebooks for code, and synthetically generated textbooks; all in English language or software code. The full list of pre-training datasets is available at the following URL: <a href="https://hf.co/collections/HuggingFaceTB/smollm3-pretraining-datasets-685a7353fdc01aecde51b1d9" class="dataset-link">https://hf.co/collections/HuggingFaceTB/smollm3-pretraining-datasets-685a7353fdc01aecde51b1d9</a></li>
209
- </ul>
210
- </li>
211
- </ul>
212
  </div>
213
  """)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
214
 
215
- gr.HTML("""
216
- <div class="info-box">
217
- <div class="subsection-header">2.2. Private non-publicly available datasets obtained from third parties</div>
218
-
219
- <h4>2.2.1. Datasets commercially licensed by rightsholders or their representatives</h4>
220
- <ul>
221
- <li><strong>Have you concluded transactional commercial licensing agreement(s) with rightsholder(s) or with their representatives?</strong>
222
- <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
223
- </li>
224
- </ul>
225
-
226
- <h4>2.2.2. Private datasets obtained from other third parties</h4>
227
- <ul>
228
- <li><strong>Have you obtained private datasets from third parties that are not licensed as described in Section 2.2.1, such as data obtained from providers of private databases, or data intermediaries?</strong>
229
- <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
230
- </li>
231
- </ul>
232
- </div>
233
- """)
234
 
235
- with gr.Row():
236
- with gr.Column():
237
- gr.HTML("""
238
- <div class="info-box">
239
- <div class="subsection-header">2.3. Data crawled and scraped from online sources</div>
240
- <ul>
241
- <li><strong>Were crawlers used by the provider or on behalf of?</strong>
242
- <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
243
- </li>
244
- </ul>
245
- </div>
246
- """)
247
-
248
- with gr.Column():
249
- gr.HTML("""
250
- <div class="info-box">
251
- <div class="subsection-header">2.4. User data</div>
252
- <ul>
253
- <li><strong>Was data from user interactions with the AI model (e.g. user input and prompts) used to train the model?</strong>
254
- <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
255
- </li>
256
- <li><strong>Was data collected from user interactions with the provider's other services or products used to train the model?</strong>
257
- <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
258
- </li>
259
- </ul>
260
- </div>
261
- """)
262
 
263
- with gr.Row():
264
- with gr.Column():
265
- gr.HTML("""
266
- <div class="info-box">
267
- <div class="subsection-header">2.5. Synthetic data</div>
268
- <ul>
269
- <li><strong>Was synthetic AI-generated data created by the provider or on their behalf to train the model?</strong>
270
- <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
271
- </li>
272
- </ul>
273
- </div>
274
- """)
275
-
276
- with gr.Column():
277
- gr.HTML("""
278
- <div class="info-box">
279
- <div class="subsection-header">2.6. Other sources of data</div>
280
- <ul>
281
- <li><strong>Have data sources other than those described in Sections 2.1 to 2.5 been used to train the model?</strong>
282
- <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
283
- </li>
284
- </ul>
285
- </div>
286
- """)
287
 
288
  # Section 3: Data Processing
289
  gr.HTML('<div class="section-header">3. Data processing aspects</div>')
290
-
291
  gr.HTML("""
292
- <div class="info-box">
293
- <div class="subsection-header">3.1. Respect of reservation of rights from text and data mining exception or limitation</div>
294
- <ul>
295
- <li><strong>Describe the measures implemented before model training to respect reservations of rights from the TDM exception or limitation before and during data collection, including the opt-out protocols and solutions honoured by the provider or, as applicable, by third parties from which datasets have been obtained:</strong>
296
- <ul>
297
- <li>The training corpus for SmolLM3-3B is made up of diverse pre-existing public datasets maintained by various organizations who still have their own approach to managing the TDM exception. All crawl-based data in the datasets uses the CommonCrawl archives which comply with robots.txt. Some datasets such as the Stack v2 additionally offer general opt-out mechanisms. For each dataset, the latest publicly available version was used to ensure propagation of any rights reservation expressed to the dataset custodian.</li>
298
- </ul>
299
- </li>
300
- </ul>
301
  </div>
302
  """)
 
 
 
 
 
 
 
 
 
 
 
 
 
303
 
304
- gr.HTML("""
305
- <div class="info-box">
306
- <div class="subsection-header">3.2. Removal of illegal content</div>
307
- <ul>
308
- <li><strong>General description of measures taken:</strong>
309
- <ul>
310
- <li>Each of the component datasets leveraged is the product of a distinct curation effort by its custodians to select the most desirable content. The specific approaches can typically be found in the dataset documentation. Among other factors, most of the datasets take the approach of using classifiers to identify "highly educational" samples that lowers the likelihood of illegal content.</li>
311
- </ul>
312
- </li>
313
- </ul>
314
- </div>
315
- """)
316
 
317
 
318
  return app
 
85
  """
86
 
87
  with gr.Blocks(css=custom_css, title="SmolLM3-3B EU Data Transparency") as app:
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
88
 
89
  with gr.Column(elem_classes=["main-container"]):
90
+ # Banner section with images
 
 
91
  with gr.Row():
92
+ with gr.Column(scale=1):
93
+ try:
94
+ gr.Image("eu_flag.png", height=180, show_label=False, show_download_button=False, interactive=False, container=False)
95
+ except:
96
+ gr.HTML('<div style="height: 120px;"></div>') # Placeholder if image not found
97
+
98
+ with gr.Column(scale=1):
99
  gr.HTML("""
100
+ <div style="text-align: center; padding: 20px;">
101
+ <h1 style="color: #2c3e50; margin: 0; font-size: 3em !important;">SmolLM3-3B</h1>
102
+ <h2 style="color: #667eea; margin: 10px 0 0 0; font-size: 1.5em !important;">Public Summary of Training Content</h2>
 
 
 
 
 
 
 
103
  </div>
104
  """)
105
+
106
+ with gr.Column(scale=1):
107
+ try:
108
+ gr.Image("banner.png", height=180, show_label=False, show_download_button=False, interactive=False, container=False)
109
+ except:
110
+ gr.HTML('<div style="height: 120px;"></div>') # Placeholder if image not found
111
 
112
+ gr.HTML("""
113
+ <div style="text-align: center; margin-top: 40px; padding: 20px; background: #f8f9fa; border-radius: 10px;">
114
+ <p style="color: #6c757d; margin: 0;">
115
+ This Space contains the transparency report for the <a href="https://huggingface.co/HuggingFaceTB/SmolLM3-3B">SmolLM3-3B</a> GPAI model developped by <a href="https://huggingface.co/">Hugging Face</a> following the guidelines provided by the AI Office.<br/>
116
+ For more information, see the <a href="https://digital-strategy.ec.europa.eu/en/library/explanatory-notice-and-template-public-summary-training-content-general-purpose-ai-models" class="dataset-link">Explanatory Notice and Template</a>
117
+ </p>
118
+ </div>
119
+ <div style="margin: 30px 0; padding: 20px; background: linear-gradient(90deg, #e3f2fd 0%, #f3e5f5 100%); border-radius: 10px; border-left: 5px solid #667eea;">
120
+ <h3 style="color: #2c3e50; margin-top: 0; font-size: 1.3em !important;"><strong>πŸ“‹ TL;DR</strong></h3>
121
+ <p style="font-size: 16px !important; line-height: 1.6; margin: 10px 0;"><strong>SmolLM3-3B</strong> is a state-of-the-art 3-billion parameter language model by <strong>Hugging Face</strong> trained on <strong>10+ trillion tokens</strong> from publicly available datasets including web documents, scientific articles, and code.
122
+ Training focused on <strong>6 EU languages</strong> plus others. The model uses <strong>only public datasets</strong> (no commercial licensing, user data, or synthetic data).
123
+ Data processing was done by the original component dataset curators with <strong>varied approaches to TDM and filtering</strong> that typically include compliance with robots.txt and other opt-out mechanisms, and educational content classifiers.</p>
124
+ </div>
125
+ """)
126
 
127
+ # Section 1: General Information
128
+ gr.HTML('<div class="section-header">1. General information</div>')
129
  gr.HTML("""
130
+ <div style="padding: 15px; margin: 10px 0; background: #f8f9fa; border-radius: 8px; border-left: 4px solid #667eea;">
131
+ <p style="margin: 0; font-size: 16px !important; color: #2c3e50;"><strong>TL;DR:</strong> Provider: Hugging Face | Model: SmolLM3-3B | Training: 10+ trillion tokens, 6 EU languages + others</p>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
132
  </div>
133
  """)
134
+ with gr.Accordion("πŸ‘‡ Click for full information", open=False):
135
+ with gr.Row():
136
+ with gr.Column():
137
+ gr.HTML("""
138
+ <div class="info-box">
139
+ <div class="subsection-header">1.1. Provider identification</div>
140
+ <ul>
141
+ <li><strong>Provider name and contact details:</strong>
142
+ <ul>
143
+ <li><strong>Hugging Face</strong></li>
144
+ <li><strong>Website: <a href="https://huggingface.co" class="dataset-link">https://huggingface.co</a></strong></li>
145
+ </ul>
146
+ </li>
147
+ </ul>
148
+ </div>
149
+ """)
150
+
151
+ with gr.Column():
152
+ gr.HTML("""
153
+ <div class="info-box">
154
+ <div class="subsection-header">1.2. Model identification</div>
155
+ <ul>
156
+ <li><strong>Versioned model name(s):</strong>
157
+ <ul><li><strong>SmolLM3-3B</strong></li></ul>
158
+ </li>
159
+ <li><strong>Model dependencies:</strong>
160
+ <ul><li><strong>None</strong></li></ul>
161
+ </li>
162
+ </ul>
163
+ </div>
164
+ """)
165
+
166
+ gr.HTML("""
167
+ <div class="info-box">
168
+ <div class="subsection-header">1.3. Modalities, overall training data size and other characteristics</div>
169
+ <ul>
170
+ <li><strong>TEXT</strong>
171
+ <ul>
172
+ <li><strong>Size:</strong> <strong>more than 10 trillion tokens</strong></li>
173
+ <li>The training corpus for SmolLM3 is made up of several publicly accessible large datasets containing web documents, scientific articles, software code, and synthetically generated textbooks and mathematical data for pre-training in addition to several mid-training and fine-tuning datasets to enable chat interactions, instruction-following and task-solving behaviors.</li>
174
+ </ul>
175
+ </li>
176
+ <li><strong>Latest date of data acquisition/collection for model training:</strong>
177
+ <ul>
178
+ <li>The training dataset is made up of different subsets with different publication and cutoff dates. For pre-training, the earliest dataset was last updated on 4/3/2024 (Stack v2), and the latest on 2/19/2025 (FineWeb2-HQ)</li>
179
+ </ul>
180
+ </li>
181
+ <li><strong>Description of the linguistic characteristics of the overall training data:</strong>
182
+ <ul>
183
+ <li>The overall training process focuses on 6 languages that are all Union languages: English, French, Spanish, German, Italian, and Portuguese. In addition, pre-training intentionally included smaller quantities of data in Mandarin Chinese, Russian, Persian, Japanese, Korean, Vietnamese, Hindi, Thai, and Greek. Other languages may have been included due to the limitations of automatic language identification in filtering stages.</li>
184
+ </ul>
185
+ </li>
186
+ <li><strong>Other relevant characteristics of the overall training data:</strong>
187
+ <ul>
188
+ <li>The training data also includes software code in the programming languages included in the Stack v2 dataset (16 languages including C, Python, Java, Markdown, HTML, Shell, etc.).</li>
189
+ </ul>
190
+ </li>
191
+ </ul>
192
+ </div>
193
+ """)
194
 
195
  # Section 2: Data Sources
196
  gr.HTML('<div class="section-header">2. List of data sources</div>')
 
197
  gr.HTML("""
198
+ <div style="padding: 15px; margin: 10px 0; background: #f8f9fa; border-radius: 8px; border-left: 4px solid #667eea;">
199
+ <p style="margin: 0; font-size: 16px !important; color: #2c3e50;"><strong>TL;DR:</strong> βœ… Publicly available datasets (DCLM, FineWeb, Stack v2, etc.) | ❌ No commercial licensing, crawling, user data, or private synthetic data</p>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
200
  </div>
201
  """)
202
+ with gr.Accordion("πŸ‘‡ Click for full information", open=False):
203
+ gr.HTML("""
204
+ <div class="info-box">
205
+ <div class="subsection-header">2.1. Publicly available datasets</div>
206
+ <ul>
207
+ <li><strong>Have you used publicly available datasets to train the model?</strong>
208
+ <ul><li><strong><span class="checkbox-yes">β˜‘ Yes</span></strong></li></ul>
209
+ </li>
210
+ <li><strong>If yes, specify the modality(ies) of the content covered by the datasets concerned:</strong>
211
+ <ul><li><strong><span class="checkbox-yes">β˜‘ Text</span></strong></li></ul>
212
+ </li>
213
+ <li><strong>List of large publicly available datasets:</strong>
214
+ <ul>
215
+ <li>DCLM: <a href="https://hf.co/datasets/mlfoundations/dclm-baseline-1.0" class="dataset-link">https://hf.co/datasets/mlfoundations/dclm-baseline-1.0</a></li>
216
+ <li>FineWeb-Edu: <a href="https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu" class="dataset-link">https://hf.co/datasets/HuggingFaceFW/fineweb-edu</a></li>
217
+ <li>FineWeb2: <a href="https://huggingface.co/datasets/epfml/FineWeb2-HQ" class="dataset-link">https://hf.co/datasets/epfml/FineWeb2-HQ</a></li>
218
+ <li>Stack V2: <a href="https://hf.co/datasets/bigcode/the-stack-v2" class="dataset-link">https://hf.co/datasets/bigcode/the-stack-v2</a></li>
219
+ <li>pes2o: <a href="https://hf.co/datasets/allenai/peS2o" class="dataset-link">https://hf.co/datasets/allenai/peS2o</a></li>
220
+ <li>SmolTalk2: <a href="https://huggingface.co/datasets/HuggingFaceTB/smoltalk2" class="dataset-link">https://hf.co/datasets/HuggingFaceTB/smoltalk2</a></li>
221
+ </ul>
222
+ </li>
223
+ <li><strong>General description of other publicly available datasets not listed above:</strong>
224
+ <ul>
225
+ <li>In addition to the large datasets cited above, many additional publicly available datasets were added to target specific domains, including several math datasets made up of both web-filtered and synthetic data, Wikipedia data, "reasoning data" generated by selected large models on diverse problems, Jupyter notebooks for code, and synthetically generated textbooks; all in English language or software code. The full list of pre-training datasets is available at the following URL: <a href="https://hf.co/collections/HuggingFaceTB/smollm3-pretraining-datasets-685a7353fdc01aecde51b1d9" class="dataset-link">https://hf.co/collections/HuggingFaceTB/smollm3-pretraining-datasets-685a7353fdc01aecde51b1d9</a></li>
226
+ </ul>
227
+ </li>
228
+ </ul>
229
+ </div>
230
+ """)
231
 
232
+ gr.HTML("""
233
+ <div class="info-box">
234
+ <div class="subsection-header">2.2. Private non-publicly available datasets obtained from third parties</div>
235
+
236
+ <h4>2.2.1. Datasets commercially licensed by rightsholders or their representatives</h4>
237
+ <ul>
238
+ <li><strong>Have you concluded transactional commercial licensing agreement(s) with rightsholder(s) or with their representatives?</strong>
239
+ <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
240
+ </li>
241
+ </ul>
242
+
243
+ <h4>2.2.2. Private datasets obtained from other third parties</h4>
244
+ <ul>
245
+ <li><strong>Have you obtained private datasets from third parties that are not licensed as described in Section 2.2.1, such as data obtained from providers of private databases, or data intermediaries?</strong>
246
+ <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
247
+ </li>
248
+ </ul>
249
+ </div>
250
+ """)
251
 
252
+ with gr.Row():
253
+ with gr.Column():
254
+ gr.HTML("""
255
+ <div class="info-box">
256
+ <div class="subsection-header">2.3. Data crawled and scraped from online sources</div>
257
+ <ul>
258
+ <li><strong>Were crawlers used by the provider or on behalf of?</strong>
259
+ <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
260
+ </li>
261
+ </ul>
262
+ </div>
263
+ """)
264
+
265
+ with gr.Column():
266
+ gr.HTML("""
267
+ <div class="info-box">
268
+ <div class="subsection-header">2.4. User data</div>
269
+ <ul>
270
+ <li><strong>Was data from user interactions with the AI model (e.g. user input and prompts) used to train the model?</strong>
271
+ <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
272
+ </li>
273
+ <li><strong>Was data collected from user interactions with the provider's other services or products used to train the model?</strong>
274
+ <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
275
+ </li>
276
+ </ul>
277
+ </div>
278
+ """)
279
 
280
+ with gr.Row():
281
+ with gr.Column():
282
+ gr.HTML("""
283
+ <div class="info-box">
284
+ <div class="subsection-header">2.5. Synthetic data</div>
285
+ <ul>
286
+ <li><strong>Was synthetic AI-generated data created by the provider or on their behalf to train the model?</strong>
287
+ <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
288
+ </li>
289
+ </ul>
290
+ </div>
291
+ """)
292
+
293
+ with gr.Column():
294
+ gr.HTML("""
295
+ <div class="info-box">
296
+ <div class="subsection-header">2.6. Other sources of data</div>
297
+ <ul>
298
+ <li><strong>Have data sources other than those described in Sections 2.1 to 2.5 been used to train the model?</strong>
299
+ <ul><li><strong><span class="checkbox-no">☐ No</span></strong></li></ul>
300
+ </li>
301
+ </ul>
302
+ </div>
303
+ """)
304
 
305
  # Section 3: Data Processing
306
  gr.HTML('<div class="section-header">3. Data processing aspects</div>')
 
307
  gr.HTML("""
308
+ <div style="padding: 15px; margin: 10px 0; background: #f8f9fa; border-radius: 8px; border-left: 4px solid #667eea;">
309
+ <p style="margin: 0; font-size: 16px !important; color: #2c3e50;"><strong>TL;DR:</strong> TDM rights: robots.txt baseline otherwise dataset-dependent | Content filtering: Dataset-dependent including educational classifiers</p>
 
 
 
 
 
 
 
310
  </div>
311
  """)
312
+ with gr.Accordion("πŸ‘‡ Click for full information", open=False):
313
+ gr.HTML("""
314
+ <div class="info-box">
315
+ <div class="subsection-header">3.1. Respect of reservation of rights from text and data mining exception or limitation</div>
316
+ <ul>
317
+ <li><strong>Describe the measures implemented before model training to respect reservations of rights from the TDM exception or limitation before and during data collection, including the opt-out protocols and solutions honoured by the provider or, as applicable, by third parties from which datasets have been obtained:</strong>
318
+ <ul>
319
+ <li>The training corpus for SmolLM3-3B is made up of diverse pre-existing public datasets maintained by various organizations who still have their own approach to managing the TDM exception. All crawl-based data in the datasets uses the CommonCrawl archives which comply with robots.txt. Some datasets such as the Stack v2 additionally offer general opt-out mechanisms. For each dataset, the latest publicly available version was used to ensure propagation of any rights reservation expressed to the dataset custodian.</li>
320
+ </ul>
321
+ </li>
322
+ </ul>
323
+ </div>
324
+ """)
325
 
326
+ gr.HTML("""
327
+ <div class="info-box">
328
+ <div class="subsection-header">3.2. Removal of illegal content</div>
329
+ <ul>
330
+ <li><strong>General description of measures taken:</strong>
331
+ <ul>
332
+ <li>Each of the component datasets leveraged is the product of a distinct curation effort by its custodians to select the most desirable content. The specific approaches can typically be found in the dataset documentation. Among other factors, most of the datasets take the approach of using classifiers to identify "highly educational" samples that lowers the likelihood of illegal content.</li>
333
+ </ul>
334
+ </li>
335
+ </ul>
336
+ </div>
337
+ """)
338
 
339
 
340
  return app