Hugging Face's logo

yhavinga
/
t5-small-24L-ccmatrix-multi

Translation
Transformers
PyTorch
JAX
TensorBoard
Dutch
English
t5
text2text-generation
seq2seq
text-generation-inference
Inference Endpoints
Model card Files Metrics Community
2
Yeb Havinga commited on
Commit
1ca33c4
•
1 Parent(s): 612f0ae

Autoupdate README.md

Browse files
Files changed (1) hide show
  1. README.md +137 -9
README.md CHANGED
@@ -1,22 +1,150 @@
1
  ---
2
- language:
3
- - en
4
  - nl
5
  datasets:
6
  - yhavinga/mc4_nl_cleaned
7
  - yhavinga/ccmatrix
8
  tags:
 
9
  - translation
 
 
 
 
 
 
 
10
  license: apache-2.0
11
  ---
12
 
13
- ## PreTraining
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
 
15
- The model was pre-trained on a English and Dutch mC4 cleaned.
 
16
 
17
- ## Finetuning
18
 
19
- The model was finetuned on CCMatrix, validated on Tatoeba.
20
- * **128-max token length**
21
- * Only the first 25M sentences of CCMatrix were used, both en->nl and en->nl (total 50M sentences).
22
- * Note: multi-direction. Prepend either `translate Dutch to English: ` or `translate English to Dutch: `
 
1
  ---
2
+ language:
 
3
  - nl
4
  datasets:
5
  - yhavinga/mc4_nl_cleaned
6
  - yhavinga/ccmatrix
7
  tags:
8
+ - t5
9
  - translation
10
+ - seq2seq
11
+
12
+ pipeline_tag: translation
13
+ widget:
14
+ - text: "It is a painful and tragic spectacle that rises before me: I have drawn back the curtain from the rottenness of man. This word, in my mouth, is at least free from one suspicion: that it involves a moral accusation against humanity. It is used--and I wish to emphasize the fact again--without any moral significance: and this is so far true that the rottenness I speak of is most apparent to me precisely in those quarters where there has been most aspiration, hitherto, toward 'virtue' and 'godliness.'"
15
+ - text: "For once Fletcher’s sedate features showed a certain lightness. 'I believe I will linger awhile longer.' He indicated a holoscreen which was displaying the image from an external camera. Cloud-splattered landscape was rolling past, pastel greens, browns, and blues illuminated by Duke’s radiance. 'It is not often a mortal man is permitted to view a world over the shoulder of angels.'"
16
+
17
  license: apache-2.0
18
  ---
19
 
20
+ # t5-small-24L-ccmatrix-multi
21
+
22
+ A [T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.html) sequence to sequence model
23
+ pre-trained from scratch on [cleaned Dutch 🇳🇱🇧🇪 mC4](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned).
24
+
25
+
26
+ This **t5 eff** model has **249M** parameters.
27
+ It was pre-trained on the dataset
28
+ `mc4_nl_cleaned` config `large_en_nl` for **1** epoch(s) and a duration of **4d10h**,
29
+ with a sequence length of **512**, batch size **128** and **851852** total steps.
30
+ Pre-training evaluation loss and accuracy are **1,18** and **0,74**.
31
+
32
+
33
+
34
+
35
+ ## Tokenizer
36
+
37
+ The model uses a cased SentencePiece tokenizer configured with the `Nmt, NFKC, Replace multi-space to single-space` normalizers
38
+ and has 32003 tokens.
39
+ It was trained on Dutch and English with scripts from the Huggingface Transformers [Flax examples](https://github.com/huggingface/transformers/tree/master/examples/flax/language-modeling).
40
+ See [./raw/main/tokenizer.json](tokenizer.json) for details.
41
+
42
+ ## Dataset
43
+
44
+ All models listed below are trained on
45
+ [cleaned Dutch mC4](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned),
46
+ which is the original mC4, except
47
+
48
+ * 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
49
+ * Sentences with less than 3 words are removed
50
+ * Sentences with a word of more than 1000 characters are removed
51
+ * Documents with less than 5 sentences are removed
52
+ * Documents with "javascript", "lorum ipsum", "terms of use", "privacy policy", "cookie policy", "uses cookies",
53
+ "use of cookies", "use cookies", "elementen ontbreken", "deze printversie" are removed.
54
+
55
+ The Dutch and English models are trained on a 50/50% mix of Dutch mC4 and English C4.
56
+
57
+ ## Models
58
+
59
+ Three types of models have been trained. `t5-base-dutch` is the only model with an original T5 config.
60
+ The other model types t5-v1.1 and t5-eff have `gated-relu` instead of `relu` as activation function,
61
+ and trained with a drop-out of `0.0` unless training would diverge (`t5-v1.1-large-dutch-cased`).
62
+ The T5-eff models are models with mostly different numbers of layers. The table will list
63
+ the several dimensions of these models. Note that `efficient` is a misnomer for models with few layers,
64
+ e.g. `t5-xl-4L-dutch-english-cased`, that is not efficient and one of the worst models on downstream summarization.
65
+
66
+ | | t5-base-dutch | t5-v1.1-base-dutch-uncased | t5-v1.1-base-dutch-cased | t5-v1.1-large-dutch-cased | t5-v1_1-base-dutch-english-cased | t5-v1_1-base-dutch-english-cased-1024 | t5-small-24L-dutch-english | t5-xl-4L-dutch-english-cased | t5-base-36L-dutch-english-cased | t5-eff-xl-8l-dutch-english-cased | t5-eff-large-8l-dutch-english-cased |
67
+ |:------------------|:----------------|:-----------------------------|:---------------------------|:----------------------------|:-----------------------------------|:----------------------------------------|:-----------------------------|:-------------------------------|:----------------------------------|:-----------------------------------|:--------------------------------------|
68
+ | 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 |
69
+ | d_model | 768 | 768 | 768 | 1024 | 768 | 768 | 512 | 2048 | 768 | 1024 | 1024 |
70
+ | d_ff | 3072 | 2048 | 2048 | 2816 | 2048 | 2048 | 1920 | 5120 | 2560 | 16384 | 4096 |
71
+ | num_heads | 12 | 12 | 12 | 16 | 12 | 12 | 8 | 32 | 12 | 32 | 16 |
72
+ | d_kv | 64 | 64 | 64 | 64 | 64 | 64 | 64 | 64 | 64 | 128 | 64 |
73
+ | num_layers | 12 | 12 | 12 | 24 | 12 | 12 | 24 | 4 | 36 | 8 | 8 |
74
+ | num parameters | 223M | 248M | 248M | 783M | 248M | 248M | 250M | 585M | 729M | 1241M | 335M |
75
+ | feed_forward_proj | relu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu |
76
+ | dropout | 0.1 | 0.0 | 0.0 | 0.1 | 0.0 | 0.0 | 0.0 | 0.1 | 0.0 | 0.0 | 0.0 |
77
+ | 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 |
78
+ | tr. seq len | 512 | 1024 | 1024 | 512 | 512 | 1024 | 512 | 512 | 512 | 512 | 512 |
79
+ | batch size | 128 | 64 | 64 | 64 | 128 | 64 | 128 | 512 | 512 | 64 | 128 |
80
+ | total steps | 527500 | 1014525 | 1210154 | 2427498 | 2839630 | 1520k/3397024 | 851852 | 212963 | 212963 | 538k/1703705 | 851850 |
81
+ | epochs | 1 | 2 | 2 | 2 | 10 | 4 | 1 | 1 | 1 | 1 | 1 |
82
+ | duration | 2d9h | 5d5h | 6d6h | 8d13h | 11d18h | 9d1h | 4d10h | 6d1h | 17d15h | 4d 19h | 3d 23h |
83
+ | optimizer | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor |
84
+ | lr | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.009 | 0.005 | 0.005 |
85
+ | warmup | 10000.0 | 10000.0 | 10000.0 | 10000.0 | 10000.0 | 5000.0 | 20000.0 | 2500.0 | 1000.0 | 1500.0 | 1500.0 |
86
+ | eval loss | 1,38 | 1,20 | 0,96 | 1,07 | 1,11 | 1,13 | 1,18 | 1,27 | 1,05 | 1,3019 | 1,15 |
87
+ | eval acc | 0,70 | 0,73 | 0,78 | 0,76 | 0,75 | 0,74 | 0,74 | 0,72 | 0,76 | 0,71 | 0,74 |
88
+
89
+ ## Evaluation on summarization
90
+
91
+ The models below have been evaluated on the summarization downstream task on 50K samples from the CNN Dailymail dataset.
92
+ All models were fine-tuned with the AdamW optimizer with a batch size of 128 and constant learning rate of 1e-3 after a
93
+ warmup of 64 steps, with a label smoothing factor of 0.05.
94
+ Article and summary token lengths were set to 1024 and 142.
95
+
96
+ | | t5-base-dutch | t5-v1.1-base-dutch-uncased | t5-v1.1-base-dutch-cased | t5-v1_1-base-dutch-english-cased | t5-v1_1-base-dutch-english-cased-1024 | t5-small-24L-dutch-english | t5-xl-4L-dutch-english-cased | t5-base-36L-dutch-english-cased | t5-eff-large-8l-dutch-english-cased | mt5-base |
97
+ |:-------------------|:----------------|:-----------------------------|:---------------------------|:-----------------------------------|:----------------------------------------|:-----------------------------|:-------------------------------|:----------------------------------|:--------------------------------------|:-----------|
98
+ | rouge1 | 33.0313 | 33.8432 | 34.0906 | 33.1116 | 34.6465 | 34.376 | 30.8983 | 35.0931 | 33.9293 | 33.6466 |
99
+ | rouge2 | 12.9452 | 13.7706 | 13.6203 | 13.275 | 13.8525 | 13.8939 | 11.6005 | 14.3823 | 13.6274 | 13.1085 |
100
+ | rougeL | 23.7204 | 24.5642 | 24.7304 | 24.3561 | 24.721 | 25.2496 | 22.6536 | 25.3213 | 24.5595 | 23.909 |
101
+ | rougeLsum | 29.842 | 30.7783 | 31.1438 | 30.0548 | 31.6104 | 31.3838 | 27.8467 | 32.3526 | 30.952 | 30.5054 |
102
+ | gen_len | 90.488 | 91.832 | 92.122 | 89.583 | 98.333 | 90.442 | 92.342 | 96.832 | 95.057 | 96.312 |
103
+ | num parameters | 223M | 248M | 248M | 248M | 248M | 250M | 585M | 729M | 335M | 582M |
104
+ | samples_per_second | 3.195 | 3.039 | 3.0 | 3.216 | 2.974 | 1.594 | 2.47 | 0.623 | 3.087 | 1.201 |
105
+
106
+ ## Translation models
107
+
108
+ The small 24L and base 36L models have been fine-tuned for translation on the CCMatrix dataset.
109
+ The models named *-`multi` support both directions of translation. The models are trained on CCMatrix only. As this is
110
+ a really large dataset with over 100M Dutch-English sentence pairs, the models are trained on a fraction of it,
111
+ refer to the table below for how long. Evaluation is performed on a CCMatrix section not trained on, but also
112
+ on Tatoeba and Opus Books. The `_bp` columns list the *brevity penalty*. The `avg_bleu` score is the bleu score
113
+ averaged over all three evaluation datasets.
114
+
115
+ The translation metrics are listed in the table below:
116
+
117
+ | | t5-base-36L-ccmatrix-en-nl | t5-base-36L-ccmatrix-multi | t5-base-36L-ccmatrix-multi | t5-small-24L-ccmatrix-multi | t5-small-24L-ccmatrix-multi |
118
+ |:-----------------------|:-----------------------------|:-----------------------------|:-----------------------------|:------------------------------|:------------------------------|
119
+ | id | 0 | 14 | 15 | 16 | 20 |
120
+ | source_lang | en | en | nl | en | nl |
121
+ | target_lang | nl | nl | en | nl | en |
122
+ | source_prefix | translate English to Dutch: | translate English to Dutch: | translate Dutch to English: | translate English to Dutch: | translate Dutch to English: |
123
+ | tatoeba_bp | 0.9897614370103832 | 0.9736173618072754 | 0.943521164106552 | 0.9760983304454847 | 0.9406676405486575 |
124
+ | ccmatrix_bp | 0.9590750786190209 | 0.9536276245543676 | 0.9635673583308255 | 0.9517934939463099 | 0.9585648049711814 |
125
+ | opus_books_bp | 0.7478011343203491 | 0.7950194726093107 | 0.9362852511299413 | 0.770498474692027 | 0.8870675076932444 |
126
+ | tatoeba_score | 50.63006965176505 | 46.580601850286214 | 52.82030981131822 | 46.419809813946046 | 51.67887417355214 |
127
+ | ccmatrix_score | 60.33227938980884 | 56.81297258845844 | 62.836646082246254 | 57.404319674892406 | 63.08633155239932 |
128
+ | opus_books_score | 10.405013868050663 | 13.477997378535864 | 24.93113308798125 | 12.927244801365507 | 23.418552148252047 |
129
+ | avg_bleu | 40.455787636541515 | 38.95719060576017 | 46.86269632718191 | 38.91712476340132 | 46.0612526247345 |
130
+ | total steps | 78125 | 390625 | 390625 | 390625 | 390625 |
131
+ | duration | 14h | 101h | 101h | 74h | 74h |
132
+ | num_parameters | 728928000 | 728928000 | 728928000 | 249991680 | 249991680 |
133
+ | label_smoothing_factor | 0.09 | 0.15 | 0.15 | 0.1 | 0.1 |
134
+ | learning_rate | 0.0001 | 5e-05 | 5e-05 | 0.0005 | 0.0005 |
135
+
136
+ ## Acknowledgements
137
+
138
+ This project would not have been possible without compute generously provided by Google through the
139
+ [TPU Research Cloud](https://sites.research.google/trc/). The HuggingFace 🤗 ecosystem and was also
140
+ instrumental all parts of the training. Logging metrics to Weights & Biases made it possible to keep track of many
141
+ models and orchestrate hyper-parameter sweeps with insightful visualizations. I cannot imagine how I would
142
+ have completed this project otherwise.
143
+ The following repositories where helpful in setting up the TPU-VM,
144
+ and getting an idea what sensible hyper-parameters are for training gpt2 from scratch.
145
 
146
+ * [Gsarti's Pretrain and Fine-tune a T5 model with Flax on GCP](https://github.com/gsarti/t5-flax-gcp)
147
+ * [Flax/Jax Community week t5-base-dutch](https://huggingface.co/flax-community/t5-base-dutch)
148
 
149
+ Created by [Yeb Havinga](https://www.linkedin.com/in/yeb-havinga-86530825/)
150