README.md CHANGED
@@ -6,35 +6,32 @@ language:
6
  - pt
7
  tags:
8
  - falcon3
9
- license: other
10
- license_name: falcon-llm-license
11
  license_link: https://falconllm.tii.ae/falcon-terms-and-conditions.html
12
- library_name: transformers
13
  ---
14
 
15
- <div align="center">
16
- <img src="https://huggingface.co/datasets/tiiuae/documentation-images/resolve/main/general/falco3-logo.png" alt="drawing" width="500"/>
17
- </div>
18
 
19
  # Falcon3-7B-Base
20
 
21
  **Falcon3** family of Open Foundation Models is a set of pretrained and instruct LLMs ranging from 1B to 10B.
22
 
23
- This repository contains the **Falcon3-7B-Base**. It achieves state of art results (at the time of release) on reasoning, language understanding, instruction following, code and mathematics tasks.
24
- Falcon3-7B-Base supports 4 languages (english, french, spanish, portuguese) and a context length up to 32K.
25
 
26
- ⚠️ **This is a raw, pretrained model, which should be further finetuned for most usecases.**
27
 
28
  ## Model Details
29
  - Architecture
30
- - transformer based causal decoder only architecture
31
  - 28 decoder blocks
32
- - grouped query attention (GQA) for faster inference: 12 query heads and 4 KV heads
33
- - wider head dimension: 256
34
- - high RoPE value to support long context understanding: 1000042
35
- - 32k context length
36
- - 131k vocab size
37
- - Pretrained on 14 Teratokens of datasets comprising of web, code, STEM, high quality and mutlilingual data using 1024 H100 GPU chips
 
38
  - Supports EN, FR, ES, PT
39
  - Developed by [Technology Innovation Institute](https://www.tii.ae)
40
  - License: TII Falcon-LLM License 2.0
@@ -65,10 +62,7 @@ print(response[0]['generated_text'])
65
  <br>
66
 
67
  ## Benchmarks
68
- We report in the following table our internal pipeline benchmarks.
69
- - We use [lm-evaluation harness](https://github.com/EleutherAI/lm-evaluation-harness).
70
- - We report **raw scores**.
71
- - We use same batch-size across all models.
72
 
73
 
74
 
@@ -79,7 +73,6 @@ We report in the following table our internal pipeline benchmarks.
79
  <col style="width: 7%;">
80
  <col style="width: 7%;">
81
  <col style="width: 7%;">
82
- <col style="width: 7%;">
83
  <col style="background-color: rgba(80, 15, 213, 0.5); width: 7%;">
84
  </colgroup>
85
  <thead>
@@ -87,7 +80,6 @@ We report in the following table our internal pipeline benchmarks.
87
  <th>Category</th>
88
  <th>Benchmark</th>
89
  <th>Llama3.1-8B</th>
90
- <th>Qwen2-7B</th>
91
  <th>Qwen2.5-7B</th>
92
  <th>gemma-2-9b</th>
93
  <th>Falcon3-7B-Base</th>
@@ -98,119 +90,101 @@ We report in the following table our internal pipeline benchmarks.
98
  <td rowspan="3">General</td>
99
  <td>MMLU (5-shot)</td>
100
  <td>65.2</td>
101
- <td>70.4</td>
102
- <td>74.2</td>
103
- <td>-</td>
104
  <td>67.5</td>
105
  </tr>
106
  <tr>
107
  <td>MMLU-PRO (5-shot)</td>
108
  <td>32.7</td>
109
- <td>42.1</td>
110
- <td>43.5</td>
111
- <td>-</td>
112
  <td>39.2</td>
113
  </tr>
114
  <tr>
115
  <td>IFEval</td>
116
  <td>12.0</td>
117
- <td>30.6</td>
118
  <td>33.9</td>
119
- <td>-</td>
120
- <td>34.3</td>
121
  </tr>
122
  <tr>
123
  <td rowspan="2">Math</td>
124
  <td>GSM8K (5-shot)</td>
125
  <td>49.4</td>
126
- <td>77.9</td>
127
- <td>82.9</td>
128
- <td>-</td>
129
  <td>76.2</td>
130
  </tr>
131
  <tr>
132
- <td>MATH(4-shot)</td>
133
  <td>4.1</td>
134
- <td>17.5</td>
135
  <td>15.5</td>
136
- <td>-</td>
137
- <td>18.0</td>
138
  </tr>
139
  <tr>
140
  <td rowspan="4">Reasoning</td>
141
  <td>Arc Challenge (25-shot)</td>
142
- <td>53.4</td>
143
- <td>57.4</td>
144
- <td>59.0</td>
145
- <td>-</td>
146
- <td>59.6</td>
147
  </tr>
148
  <tr>
149
  <td>GPQA (0-shot)</td>
150
  <td>31.0</td>
151
- <td>31.9</td>
152
  <td>33.0</td>
153
- <td>-</td>
154
- <td>35.5</td>
155
  </tr>
156
  <tr>
157
  <td>MUSR (0-shot)</td>
158
  <td>38.0</td>
159
- <td>44.1</td>
160
  <td>44.2</td>
161
- <td>-</td>
162
- <td>47.3</td>
163
  </tr>
164
  <tr>
165
  <td>BBH (3-shot)</td>
166
  <td>46.5</td>
167
- <td>53.3</td>
168
  <td>54.0</td>
169
- <td>-</td>
170
  <td>51.0</td>
171
  </tr>
172
  <tr>
173
  <td rowspan="4">CommonSense Understanding</td>
174
  <td>PIQA (0-shot)</td>
175
- <td>80.3</td>
176
- <td>79.8</td>
177
- <td>78.7</td>
178
- <td>-</td>
179
- <td>77.7</td>
180
  </tr>
181
  <tr>
182
  <td>SciQ (0-shot)</td>
183
- <td>96.3</td>
184
- <td>95.9</td>
185
- <td>96.6</td>
186
- <td>-</td>
187
- <td>95.3</td>
188
  </tr>
189
  <tr>
190
  <td>Winogrande (0-shot)</td>
191
  <td>74.0</td>
192
- <td>72.1</td>
193
  <td>72.9</td>
194
- <td>-</td>
195
  <td>71.0</td>
196
  </tr>
197
  <tr>
198
  <td>OpenbookQA (0-shot)</td>
199
- <td>33.4</td>
200
- <td>35.2</td>
201
- <td>33.6</td>
202
- <td>-</td>
203
- <td>31.4</td>
204
  </tr>
205
  </tbody>
206
  </table>
207
 
208
- ## Useful links
209
- - View our [release blogpost](https://huggingface.co/blog/falcon3).
210
- - Feel free to join [our discord server](https://discord.gg/fwXpMyGc) if you have any questions or to interact with our researchers and developers.
211
-
212
  ## Technical Report
213
-
214
  Coming soon....
215
 
216
  ## Citation
@@ -218,7 +192,7 @@ If Falcon3 family were helpful to your work, feel free to give us a cite.
218
 
219
  ```
220
  @misc{Falcon3,
221
- title = {Falcon 3 family of Open Foundation Models},
222
  author = {TII Team},
223
  month = {December},
224
  year = {2024}
 
6
  - pt
7
  tags:
8
  - falcon3
9
+ license: other
10
+ license_name: falcon-llm-license
11
  license_link: https://falconllm.tii.ae/falcon-terms-and-conditions.html
 
12
  ---
13
 
 
 
 
14
 
15
  # Falcon3-7B-Base
16
 
17
  **Falcon3** family of Open Foundation Models is a set of pretrained and instruct LLMs ranging from 1B to 10B.
18
 
19
+ This repository contains the **Falcon3-7B-Base**. It achieves state-of-the-art results (at release's time) on reasoning, language understanding, instruction following, code and mathematics tasks.
20
+ Falcon3-7B-Base supports 4 languages (English, French, Spanish, Portuguese) and a context length of up to 32K.
21
 
22
+ ⚠️ **This is a raw, pretrained model, which should be further finetuned using SFT, RLHF, continued pretraining, etc. for most use cases.**
23
 
24
  ## Model Details
25
  - Architecture
26
+ - Transformer-based causal decoder-only architecture
27
  - 28 decoder blocks
28
+ - Grouped Query Attention (GQA) for faster inference: 12 query heads and 4 key-value heads
29
+ - Wider head dimension: 256
30
+ - High RoPE value to support long context understanding: 1000042
31
+ - Uses SwiGLU and RMSNorm
32
+ - 32K context length
33
+ - 131K vocab size
34
+ - Pretrained on 14 Teratokens of datasets comprising of web, code, STEM, high quality and mutlilingual data using 2048 H100 GPU chips
35
  - Supports EN, FR, ES, PT
36
  - Developed by [Technology Innovation Institute](https://www.tii.ae)
37
  - License: TII Falcon-LLM License 2.0
 
62
  <br>
63
 
64
  ## Benchmarks
65
+ We report in the following table our internal pipeline benchmarks:
 
 
 
66
 
67
 
68
 
 
73
  <col style="width: 7%;">
74
  <col style="width: 7%;">
75
  <col style="width: 7%;">
 
76
  <col style="background-color: rgba(80, 15, 213, 0.5); width: 7%;">
77
  </colgroup>
78
  <thead>
 
80
  <th>Category</th>
81
  <th>Benchmark</th>
82
  <th>Llama3.1-8B</th>
 
83
  <th>Qwen2.5-7B</th>
84
  <th>gemma-2-9b</th>
85
  <th>Falcon3-7B-Base</th>
 
90
  <td rowspan="3">General</td>
91
  <td>MMLU (5-shot)</td>
92
  <td>65.2</td>
93
+ <td><b>74.2</b></td>
94
+ <td>70.8</td>
 
95
  <td>67.5</td>
96
  </tr>
97
  <tr>
98
  <td>MMLU-PRO (5-shot)</td>
99
  <td>32.7</td>
100
+ <td><b>43.5</b></td>
101
+ <td>41.4</td>
 
102
  <td>39.2</td>
103
  </tr>
104
  <tr>
105
  <td>IFEval</td>
106
  <td>12.0</td>
 
107
  <td>33.9</td>
108
+ <td>21.2</td>
109
+ <td><b>34.3</b></td>
110
  </tr>
111
  <tr>
112
  <td rowspan="2">Math</td>
113
  <td>GSM8K (5-shot)</td>
114
  <td>49.4</td>
115
+ <td><b>82.9</b></td>
116
+ <td>69.1</td>
 
117
  <td>76.2</td>
118
  </tr>
119
  <tr>
120
+ <td>MATH Lvl-5 (4-shot)</td>
121
  <td>4.1</td>
 
122
  <td>15.5</td>
123
+ <td>10.5</td>
124
+ <td><b>18.0</b></td>
125
  </tr>
126
  <tr>
127
  <td rowspan="4">Reasoning</td>
128
  <td>Arc Challenge (25-shot)</td>
129
+ <td>58.2</td>
130
+ <td>63.2</td>
131
+ <td><b>67.5</b></td>
132
+ <td>63.1</td>
 
133
  </tr>
134
  <tr>
135
  <td>GPQA (0-shot)</td>
136
  <td>31.0</td>
 
137
  <td>33.0</td>
138
+ <td>33.4</td>
139
+ <td><b>35.5</b></td>
140
  </tr>
141
  <tr>
142
  <td>MUSR (0-shot)</td>
143
  <td>38.0</td>
 
144
  <td>44.2</td>
145
+ <td>45.3</td>
146
+ <td><b>47.3</b></td>
147
  </tr>
148
  <tr>
149
  <td>BBH (3-shot)</td>
150
  <td>46.5</td>
 
151
  <td>54.0</td>
152
+ <td><b>54.3</b></td>
153
  <td>51.0</td>
154
  </tr>
155
  <tr>
156
  <td rowspan="4">CommonSense Understanding</td>
157
  <td>PIQA (0-shot)</td>
158
+ <td>81.2</td>
159
+ <td>79.9</td>
160
+ <td><b>82.9</b></td>
161
+ <td>79.1</td>
 
162
  </tr>
163
  <tr>
164
  <td>SciQ (0-shot)</td>
165
+ <td>94.6</td>
166
+ <td>95.2</td>
167
+ <td><b>97.1</b></td>
168
+ <td>92.4</td>
 
169
  </tr>
170
  <tr>
171
  <td>Winogrande (0-shot)</td>
172
  <td>74.0</td>
 
173
  <td>72.9</td>
174
+ <td><b>74.2</b></td>
175
  <td>71.0</td>
176
  </tr>
177
  <tr>
178
  <td>OpenbookQA (0-shot)</td>
179
+ <td>44.8</td>
180
+ <td>47.0</td>
181
+ <td><b>47.2</b></td>
182
+ <td>43.8</td>
 
183
  </tr>
184
  </tbody>
185
  </table>
186
 
 
 
 
 
187
  ## Technical Report
 
188
  Coming soon....
189
 
190
  ## Citation
 
192
 
193
  ```
194
  @misc{Falcon3,
195
+ title = {The Falcon 3 family of Open Models},
196
  author = {TII Team},
197
  month = {December},
198
  year = {2024}
generation_config.json CHANGED
@@ -1,6 +1,5 @@
1
  {
2
  "_from_model_config": true,
3
- "bos_token_id": 11,
4
  "eos_token_id": 11,
5
  "transformers_version": "4.46.1"
6
  }
 
1
  {
2
  "_from_model_config": true,
 
3
  "eos_token_id": 11,
4
  "transformers_version": "4.46.1"
5
  }
special_tokens_map.json CHANGED
@@ -30,5 +30,12 @@
30
  "normalized": false,
31
  "rstrip": false,
32
  "single_word": false
 
 
 
 
 
 
 
33
  }
34
  }
 
30
  "normalized": false,
31
  "rstrip": false,
32
  "single_word": false
33
+ },
34
+ "pad_token": {
35
+ "content": "<|pad|>",
36
+ "lstrip": false,
37
+ "normalized": false,
38
+ "rstrip": false,
39
+ "single_word": false
40
  }
41
  }
tokenizer.json CHANGED
@@ -18212,7 +18212,7 @@
18212
  },
18213
  {
18214
  "id": 2023,
18215
- "content": ">>UNUSED_1897<<",
18216
  "single_word": false,
18217
  "lstrip": false,
18218
  "rstrip": false,
@@ -20280,7 +20280,7 @@
20280
  ">>UNUSED_1894<<": 2020,
20281
  ">>UNUSED_1895<<": 2021,
20282
  ">>UNUSED_1896<<": 2022,
20283
- ">>UNUSED_1897<<": 2023,
20284
  "!": 2024,
20285
  "\"": 2025,
20286
  "#": 2026,
 
18212
  },
18213
  {
18214
  "id": 2023,
18215
+ "content": "<|pad|>",
18216
  "single_word": false,
18217
  "lstrip": false,
18218
  "rstrip": false,
 
20280
  ">>UNUSED_1894<<": 2020,
20281
  ">>UNUSED_1895<<": 2021,
20282
  ">>UNUSED_1896<<": 2022,
20283
+ "<|pad|>": 2023,
20284
  "!": 2024,
20285
  "\"": 2025,
20286
  "#": 2026,
tokenizer_config.json CHANGED
@@ -16186,7 +16186,7 @@
16186
  "special": true
16187
  },
16188
  "2023": {
16189
- "content": ">>UNUSED_1897<<",
16190
  "lstrip": false,
16191
  "normalized": false,
16192
  "rstrip": false,
@@ -16221,6 +16221,12 @@
16221
  ],
16222
  "clean_up_tokenization_spaces": true,
16223
  "eos_token": "<|endoftext|>",
 
 
 
 
 
16224
  "model_max_length": 32768,
 
16225
  "tokenizer_class": "PreTrainedTokenizerFast"
16226
  }
 
16186
  "special": true
16187
  },
16188
  "2023": {
16189
+ "content": "<|pad|>",
16190
  "lstrip": false,
16191
  "normalized": false,
16192
  "rstrip": false,
 
16221
  ],
16222
  "clean_up_tokenization_spaces": true,
16223
  "eos_token": "<|endoftext|>",
16224
+ "extra_special_tokens": {},
16225
+ "model_input_names": [
16226
+ "input_ids",
16227
+ "attention_mask"
16228
+ ],
16229
  "model_max_length": 32768,
16230
+ "pad_token": "<|pad|>",
16231
  "tokenizer_class": "PreTrainedTokenizerFast"
16232
  }