File size: 27,663 Bytes
fceee51
 
 
 
 
 
b1bd3b1
fceee51
b1bd3b1
fceee51
b1bd3b1
fceee51
 
b1bd3b1
 
 
 
 
 
 
 
7efb198
b1bd3b1
 
 
 
 
a935048
 
b1bd3b1
a935048
b1bd3b1
a935048
b1bd3b1
a935048
fa6f0d1
 
 
b1bd3b1
 
 
 
 
 
7efb198
b1bd3b1
de550c0
b1bd3b1
 
 
 
 
a935048
de550c0
a935048
de550c0
 
 
 
 
 
 
 
 
 
a935048
 
 
de550c0
a935048
de550c0
a935048
 
b1bd3b1
 
a935048
 
 
b1bd3b1
a935048
b1bd3b1
a935048
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b1bd3b1
 
 
a935048
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
de550c0
 
 
 
a935048
 
 
b1bd3b1
a935048
de550c0
 
 
 
b1bd3b1
 
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
---
language:
- nl
datasets:
- yhavinga/mc4_nl_cleaned
tags:
- t5
- seq2seq

inference: false
license: apache-2.0
---

# t5-base-dutch


Created by [Yeb Havinga](https://www.linkedin.com/in/yeb-havinga-86530825/)
& [Dat Nguyen](https://www.linkedin.com/in/dat-nguyen-49a641138/) during the [Hugging Face community week](https://discuss.huggingface.co/t/open-to-the-community-community-week-using-jax-flax-for-nlp-cv/7104), organized by [HuggingFace](https://huggingface.co/) and TPU usage sponsored by Google, for the project [Pre-train T5 from scratch in Dutch](https://discuss.huggingface.co/t/pretrain-t5-from-scratch-in-dutch/8109).
See also the fine-tuned [t5-base-dutch-demo](https://huggingface.co/flax-community/t5-base-dutch-demo) model,
and the demo application **[Netherformer 📰](https://huggingface.co/spaces/flax-community/netherformer)**,
that are based on this model.

**5 jan 2022: Model updated. Evaluation accuracy increased from 0.64 to 0.70.**

**11 jan 2022: See also [yhavinga/t5-v1.1-base-dutch-cased](https://huggingface.co/yhavinga/t5-v1.1-base-dutch-cased) with eval acc 0.78**




This **t5** model has **222M** parameters.
It was pre-trained with masked language modeling (denoise token span corruption) objective on the dataset
`mc4_nl_cleaned` config `full` for **1** epoch(s) and a duration of **2d9h**,
with a sequence length of **512**, batch size **128** and **527500** total steps (**35B** tokens).
Pre-training evaluation loss and accuracy are **1,38** and **0,70**.
Refer to the evaluation section below for a comparison of the pre-trained models on summarization and translation.
* Pre-trained T5 models need to be finetuned before they can be used for downstream tasks, therefore the inference widget on the right has been turned off.
* For a demo of the Dutch CNN summarization models, head over to the Hugging Face Spaces for
the **[Netherformer 📰](https://huggingface.co/spaces/flax-community/netherformer)** example application!

Please refer to the original T5 papers and Scale Efficiently papers for more information about the T5 architecture
and configs, though it must be noted that this model (t5-base-dutch) is unrelated to these projects and not an 'official' checkpoint.
* **[Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/pdf/1910.10683.pdf)** by *Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu*.
* **[Scale Efficiently: Insights from Pre-training and Fine-tuning Transformers](https://arxiv.org/abs/2109.10686)** by *Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, Donald Metzler*.


## Tokenizer

The model uses a cased SentencePiece tokenizer configured with the `Nmt, NFKC, Replace multi-space to single-space` normalizers
and has 32003 tokens.
It was trained on Dutch mc4 with scripts from the Huggingface Transformers [Flax examples](https://github.com/huggingface/transformers/tree/master/examples/flax/language-modeling).
See [./raw/main/tokenizer.json](tokenizer.json) for details.
 
## Dataset(s)

All models listed below are pre-trained on 
[cleaned Dutch mC4](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned),
which is the original mC4, except

  * Documents that contained words from a selection of the Dutch and English [List of Dirty Naught Obscene and Otherwise Bad Words](https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words) are removed
  * Sentences with less than 3 words are removed
  * Sentences with a word of more than 1000 characters are removed
  * Documents with less than 5 sentences are removed
  * Documents with "javascript", "lorum ipsum", "terms of use", "privacy policy", "cookie policy", "uses cookies",
    "use of cookies", "use cookies", "elementen ontbreken", "deze printversie" are removed.
 
The Dutch and English models are pre-trained on a 50/50% mix of Dutch mC4 and English C4.

The translation models are fine-tuned on [CCMatrix](https://huggingface.co/datasets/yhavinga/ccmatrix).

## Dutch T5 Models

Three types of [Dutch T5 models have been trained (blog)](https://huggingface.co/spaces/yhavinga/pre-training-dutch-t5-models).
`t5-base-dutch` is the only model with an original T5 config.
The other model types t5-v1.1 and t5-eff have `gated-relu` instead of `relu` as activation function,
and trained with a drop-out of `0.0` unless training would diverge (`t5-v1.1-large-dutch-cased`).
The T5-eff models are models that differ in their number of layers. The table will list
the several dimensions of these models. Not all t5-eff models are efficient, the best example being the inefficient
`t5-xl-4L-dutch-english-cased`. 

|                   | [t5-base-dutch](https://huggingface.co/yhavinga/t5-base-dutch)   | [t5-v1.1-base-dutch-uncased](https://huggingface.co/yhavinga/t5-v1.1-base-dutch-uncased)   | [t5-v1.1-base-dutch-cased](https://huggingface.co/yhavinga/t5-v1.1-base-dutch-cased)   | [t5-v1.1-large-dutch-cased](https://huggingface.co/yhavinga/t5-v1.1-large-dutch-cased)   | [t5-v1_1-base-dutch-english-cased](https://huggingface.co/yhavinga/t5-v1_1-base-dutch-english-cased)   | [t5-v1_1-base-dutch-english-cased-1024](https://huggingface.co/yhavinga/t5-v1_1-base-dutch-english-cased-1024)   | [t5-small-24L-dutch-english](https://huggingface.co/yhavinga/t5-small-24L-dutch-english)   | [t5-xl-4L-dutch-english-cased](https://huggingface.co/yhavinga/t5-xl-4L-dutch-english-cased)   | [t5-base-36L-dutch-english-cased](https://huggingface.co/yhavinga/t5-base-36L-dutch-english-cased)   | [t5-eff-xl-8l-dutch-english-cased](https://huggingface.co/yhavinga/t5-eff-xl-8l-dutch-english-cased)   | [t5-eff-large-8l-dutch-english-cased](https://huggingface.co/yhavinga/t5-eff-large-8l-dutch-english-cased)   |
|:------------------|:----------------|:-----------------------------|:---------------------------|:----------------------------|:-----------------------------------|:----------------------------------------|:-----------------------------|:-------------------------------|:----------------------------------|:-----------------------------------|:--------------------------------------|
| *type* | t5              | t5-v1.1                      | t5-v1.1                    | t5-v1.1                     | t5-v1.1                            | t5-v1.1                                 | t5 eff                       | t5 eff                         | t5 eff                            | t5 eff                             | t5 eff                                |
| *d_model* | 768             | 768                          | 768                        | 1024                        | 768                                | 768                                     | 512                          | 2048                           | 768                               | 1024                               | 1024                                  |
| *d_ff* | 3072            | 2048                         | 2048                       | 2816                        | 2048                               | 2048                                    | 1920                         | 5120                           | 2560                              | 16384                              | 4096                                  |
| *num_heads* | 12              | 12                           | 12                         | 16                          | 12                                 | 12                                      | 8                            | 32                             | 12                                | 32                                 | 16                                    |
| *d_kv* | 64              | 64                           | 64                         | 64                          | 64                                 | 64                                      | 64                           | 64                             | 64                                | 128                                | 64                                    |
| *num_layers* | 12              | 12                           | 12                         | 24                          | 12                                 | 12                                      | 24                           | 4                              | 36                                | 8                                  | 8                                     |
| *num parameters* | 223M            | 248M                         | 248M                       | 783M                        | 248M                               | 248M                                    | 250M                         | 585M                           | 729M                              | 1241M                              | 335M                                  |
| *feed_forward_proj* | relu            | gated-gelu                   | gated-gelu                 | gated-gelu                  | gated-gelu                         | gated-gelu                              | gated-gelu                   | gated-gelu                     | gated-gelu                        | gated-gelu                         | gated-gelu                            |
| *dropout* | 0.1             | 0.0                          | 0.0                        | 0.1                         | 0.0                                | 0.0                                     | 0.0                          | 0.1                            | 0.0                               | 0.0                                | 0.0                                   |
| *dataset* | mc4_nl_cleaned  | mc4_nl_cleaned full          | mc4_nl_cleaned full        | mc4_nl_cleaned              | mc4_nl_cleaned small_en_nl         | mc4_nl_cleaned large_en_nl              | mc4_nl_cleaned large_en_nl   | mc4_nl_cleaned large_en_nl     | mc4_nl_cleaned large_en_nl        | mc4_nl_cleaned large_en_nl         | mc4_nl_cleaned large_en_nl            |
| *tr. seq len* | 512             | 1024                         | 1024                       | 512                         | 512                                | 1024                                    | 512                          | 512                            | 512                               | 512                                | 512                                   |
| *batch size* | 128             | 64                           | 64                         | 64                          | 128                                | 64                                      | 128                          | 512                            | 512                               | 64                                 | 128                                   |
| *total steps* | 527500          | 1014525                      | 1210154                    | 1120k/2427498               | 2839630                            | 1520k/3397024                           | 851852                       | 212963                         | 212963                            | 538k/1703705                       | 851850                                |
| *epochs* | 1               | 2                            | 2                          | 2                           | 10                                 | 4                                       | 1                            | 1                              | 1                                 | 1                                  | 1                                     |
| *duration* | 2d9h            | 5d5h                         | 6d6h                       | 8d13h                       | 11d18h                             | 9d1h                                    | 4d10h                        | 6d1h                           | 17d15h                            | 4d 19h                             | 3d 23h                                |
| *optimizer* | adafactor       | adafactor                    | adafactor                  | adafactor                   | adafactor                          | adafactor                               | adafactor                    | adafactor                      | adafactor                         | adafactor                          | adafactor                             |
| *lr* | 0.005           | 0.005                        | 0.005                      | 0.005                       | 0.005                              | 0.005                                   | 0.005                        | 0.005                          | 0.009                             | 0.005                              | 0.005                                 |
| *warmup* | 10000.0         | 10000.0                      | 10000.0                    | 10000.0                     | 10000.0                            | 5000.0                                  | 20000.0                      | 2500.0                         | 1000.0                            | 1500.0                             | 1500.0                                |
| *eval loss* | 1,38            | 1,20                         | 0,96                       | 1,07                        | 1,11                               | 1,13                                    | 1,18                         | 1,27                           | 1,05                              | 1,3019                             | 1,15                                  |
| *eval acc* | 0,70            | 0,73                         | 0,78                       | 0,76                        | 0,75                               | 0,74                                    | 0,74                         | 0,72                           | 0,76                              | 0,71                               | 0,74                                  |

## Evaluation

Most models from the list above have been fine-tuned for summarization and translation.
The figure below shows the evaluation scores, where the x-axis shows the translation Bleu score (higher is better)
and y-axis the summarization Rouge1 translation score (higher is better).
Point size is proportional to the model size. Models with faster inference speed are green, slower inference speed is
plotted as bleu.

![Evaluation T5 Dutch English](evaluation_t5_dutch_english.png)

Evaluation was run on fine-tuned models trained with the following settings:   


|                | Summarization    | Translation       |
|---------------:|------------------|-------------------|
|        Dataset | CNN Dailymail NL | CCMatrix en -> nl |
| #train samples | 50K              | 50K               |
|      Optimizer | Adam             | Adam              |
|  learning rate | 0.001            | 0.0005            |
|  source length | 1024             | 128               |
|  target length | 142              | 128               |
|label smoothing | 0.05             | 0.1               |
|  #eval samples | 1000             | 1000              |

Note that the amount of training data is limited to a fraction of the total dataset sizes, therefore the scores
below can only be used to compare the 'transfer-learning' strength. The fine-tuned checkpoints for this evaluation 
are not saved, since they were trained for comparison of pre-trained models only.

The numbers for summarization are the Rouge scores on 1000 documents from the test split.

|                         |   [t5-base-dutch](https://huggingface.co/yhavinga/t5-base-dutch) |   [t5-v1.1-base-dutch-uncased](https://huggingface.co/yhavinga/t5-v1.1-base-dutch-uncased) |   [t5-v1.1-base-dutch-cased](https://huggingface.co/yhavinga/t5-v1.1-base-dutch-cased) |   [t5-v1_1-base-dutch-english-cased](https://huggingface.co/yhavinga/t5-v1_1-base-dutch-english-cased) |   [t5-v1_1-base-dutch-english-cased-1024](https://huggingface.co/yhavinga/t5-v1_1-base-dutch-english-cased-1024) |   [t5-small-24L-dutch-english](https://huggingface.co/yhavinga/t5-small-24L-dutch-english) |   [t5-xl-4L-dutch-english-cased](https://huggingface.co/yhavinga/t5-xl-4L-dutch-english-cased) |   [t5-base-36L-dutch-english-cased](https://huggingface.co/yhavinga/t5-base-36L-dutch-english-cased) |   [t5-eff-large-8l-dutch-english-cased](https://huggingface.co/yhavinga/t5-eff-large-8l-dutch-english-cased) |   mt5-base |
|:------------------------|----------------:|-----------------------------:|---------------------------:|-----------------------------------:|----------------------------------------:|-----------------------------:|-------------------------------:|----------------------------------:|--------------------------------------:|-----------:|
| *rouge1* |           33.38 |                        33.97 |                      34.39 |                              33.38 |                                   34.97 |                        34.38 |                          30.35 |                             **35.04** |                                 34.04 |      33.25 |
| *rouge2* |           13.32 |                        13.85 |                      13.98 |                              13.47 |                                   14.01 |                        13.89 |                          11.57 |                             **14.23** |                                 13.76 |      12.74 |
| *rougeL* |           24.22 |                        24.72 |                      25.1  |                              24.34 |                                   24.99 |                        **25.25** |                          22.69 |                             25.05 |                                 24.75 |      23.5  |
| *rougeLsum* |           30.23 |                        30.9  |                      31.44 |                              30.51 |                                   32.01 |                        31.38 |                          27.5  |                             **32.12** |                                 31.12 |      30.15 |
| *samples_per_second* |            3.18 |                         3.02 |                       2.99 |                               3.22 |                                    2.97 |                         1.57 |                           2.8  |                              0.61 |                                  **3.27** |       1.22 |

The models below have been evaluated for English to Dutch translation.
Note that the first four models are pre-trained on Dutch only. That they still perform adequate is probably because
the translation direction is English to Dutch.
The numbers reported are the Bleu scores on 1000 documents from the test split.

|                                |   [t5-base-dutch](https://huggingface.co/yhavinga/t5-base-dutch) |   [t5-v1.1-base-dutch-uncased](https://huggingface.co/yhavinga/t5-v1.1-base-dutch-uncased) |   [t5-v1.1-base-dutch-cased](https://huggingface.co/yhavinga/t5-v1.1-base-dutch-cased) |   [t5-v1.1-large-dutch-cased](https://huggingface.co/yhavinga/t5-v1.1-large-dutch-cased) |   [t5-v1_1-base-dutch-english-cased](https://huggingface.co/yhavinga/t5-v1_1-base-dutch-english-cased) |   [t5-v1_1-base-dutch-english-cased-1024](https://huggingface.co/yhavinga/t5-v1_1-base-dutch-english-cased-1024) |   [t5-small-24L-dutch-english](https://huggingface.co/yhavinga/t5-small-24L-dutch-english) |   [t5-xl-4L-dutch-english-cased](https://huggingface.co/yhavinga/t5-xl-4L-dutch-english-cased) |   [t5-base-36L-dutch-english-cased](https://huggingface.co/yhavinga/t5-base-36L-dutch-english-cased) |   [t5-eff-large-8l-dutch-english-cased](https://huggingface.co/yhavinga/t5-eff-large-8l-dutch-english-cased) |   mt5-base |
|:-------------------------------|----------------:|-----------------------------:|---------------------------:|----------------------------:|-----------------------------------:|----------------------------------------:|-----------------------------:|-------------------------------:|----------------------------------:|--------------------------------------:|-----------:|
| *precision_ng1* |           74.17 |                        78.09 |                      77.08 |                       72.12 |                              77.19 |                                   78.76 |                        78.59 |                          77.3  |                             **79.75** |                                 78.88 |      73.47 |
| *precision_ng2* |           52.42 |                        57.52 |                      55.31 |                       48.7  |                              55.39 |                                   58.01 |                        57.83 |                          55.27 |                             **59.89** |                                 58.27 |      50.12 |
| *precision_ng3* |           39.55 |                        45.2  |                      42.54 |                       35.54 |                              42.25 |                                   45.13 |                        45.02 |                          42.06 |                             **47.4**  |                                 45.95 |      36.59 |
| *precision_ng4* |           30.23 |                        36.04 |                      33.26 |                       26.27 |                              32.74 |                                   35.72 |                        35.41 |                          32.61 |                             **38.1**  |                                 36.91 |      27.26 |
| *bp* |            0.99 |                         0.98 |                       0.97 |                        0.98 |                               0.98 |                                    0.98 |                         0.98 |                           0.97 |                              0.98 |                                  0.98 |       0.98 |
| *score* |           45.88 |                        51.21 |                      48.31 |                       41.59 |                              48.17 |                                   51.31 |                        50.82 |                          47.83 |                             **53**    |                                 51.79 |      42.74 |
| *samples_per_second* |           **45.19** |                        45.05 |                      38.67 |                       10.12 |                              42.19 |                                   42.61 |                        12.85 |                          33.74 |                              9.07 |                                 37.86 |       9.03 |


## Translation models

The models `t5-small-24L-dutch-english` and `t5-base-36L-dutch-english` have been fine-tuned for both language
directions on the first 25M samples from CCMatrix, giving a total of 50M training samples.
Evaluation is performed on out-of-sample CCMatrix and also on Tatoeba and Opus Books.
The `_bp` columns list the *brevity penalty*. The `avg_bleu` score is the bleu score
averaged over all three evaluation datasets. The best scores displayed in bold for both translation directions.

|                        | [t5-base-36L-ccmatrix-multi](https://huggingface.co/yhavinga/t5-base-36L-ccmatrix-multi)   | [t5-base-36L-ccmatrix-multi](https://huggingface.co/yhavinga/t5-base-36L-ccmatrix-multi)   | [t5-small-24L-ccmatrix-multi](https://huggingface.co/yhavinga/t5-small-24L-ccmatrix-multi)   | [t5-small-24L-ccmatrix-multi](https://huggingface.co/yhavinga/t5-small-24L-ccmatrix-multi)   |
|:-----------------------|:-----------------------------|:-----------------------------|:------------------------------|:------------------------------|
| *source_lang* | en                           | nl                           | en                            | nl                            |
| *target_lang* | nl                           | en                           | nl                            | en                            |
| *source_prefix* | translate English to Dutch:  | translate Dutch to English:  | translate English to Dutch:   | translate Dutch to English:   |
| *ccmatrix_bleu* | **56.8**                         | 62.8                         | 57.4                          | **63.1**                          |
| *tatoeba_bleu* | **46.6**                         | **52.8**                         | 46.4                          | 51.7                          |
| *opus_books_bleu* | **13.5**                         | **24.9**                         | 12.9                          | 23.4                          |
| *ccmatrix_bp* | 0.95                         | 0.96                         | 0.95                          | 0.96                          |
| *tatoeba_bp* | 0.97                         | 0.94                         | 0.98                          | 0.94                          |
| *opus_books_bp* | 0.8                          | 0.94                         | 0.77                          | 0.89                          |
| *avg_bleu* | **38.96**                        | **46.86**                        | 38.92                         | 46.06                         |
| *max_source_length* | 128                          | 128                          | 128                           | 128                           |
| *max_target_length* | 128                          | 128                          | 128                           | 128                           |
| *adam_beta1* | 0.9                          | 0.9                          | 0.9                           | 0.9                           |
| *adam_beta2* | 0.997                        | 0.997                        | 0.997                         | 0.997                         |
| *weight_decay* | 0.05                         | 0.05                         | 0.002                         | 0.002                         |
| *lr* | 5e-05                        | 5e-05                        | 0.0005                        | 0.0005                        |
| *label_smoothing_factor* | 0.15                         | 0.15                         | 0.1                           | 0.1                           |
| *train_batch_size* | 128                          | 128                          | 128                           | 128                           |
| *warmup_steps* | 2000                         | 2000                         | 2000                          | 2000                          |
| *total steps* | 390625                       | 390625                       | 390625                        | 390625                        |
| *duration* | 4d 5h                        | 4d 5h                        | 3d 2h                         | 3d 2h                         |
| *num parameters* | 729M                         | 729M                         | 250M                          | 250M                          |

## Acknowledgements

This project would not have been possible without compute generously provided by Google through the
[TPU Research Cloud](https://sites.research.google/trc/). The HuggingFace 🤗 ecosystem was instrumental in all parts
of the training. Weights & Biases made it possible to keep track of many training sessions
and orchestrate hyper-parameter sweeps with insightful visualizations.
The following repositories where helpful in setting up the TPU-VM,
and getting an idea what sensible hyper-parameters are for training gpt2 from scratch: 

* [Gsarti's Pretrain and Fine-tune a T5 model with Flax on GCP](https://github.com/gsarti/t5-flax-gcp)
* [Flax/Jax Community week t5-base-dutch](https://huggingface.co/flax-community/t5-base-dutch)

Created by [Yeb Havinga](https://www.linkedin.com/in/yeb-havinga-86530825/)