Yeb Havinga commited on
Commit
3f7f6f9
•
1 Parent(s): a2ca396

Autoupdate README.md

Browse files
Files changed (2) hide show
  1. README.md +125 -86
  2. evaluation_t5_dutch_english.png +0 -0
README.md CHANGED
@@ -17,14 +17,14 @@ A [T5](https://ai.googleblog.com/2020/02/exploring-transfer-learning-with-t5.htm
17
  pre-trained from scratch on [cleaned Dutch 🇳🇱🇧🇪 mC4](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned).
18
 
19
 
 
 
20
  This **t5-v1.1** model has **247M** parameters.
21
- It was pre-trained on the dataset
22
  `mc4_nl_cleaned` config `full` for **2** epoch(s) and a duration of **6d6h**,
23
- with a sequence length of **1024**, batch size **64** and **1210154** total steps.
24
  Pre-training evaluation loss and accuracy are **0,96** and **0,78**.
25
- After fine-tuning on 25K samples of Dutch CNN summarization, the Rouge1 score is **34.1**
26
- (note: this evaluation model was not saved).
27
-
28
  * 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.
29
  * For a demo of the Dutch CNN summarization models, head over to the Hugging Face Spaces for
30
  the **[Netherformer 📰](https://huggingface.co/spaces/flax-community/netherformer)** example application!
@@ -35,9 +35,6 @@ and configs, though it must be noted that this model (t5-v1.1-base-dutch-cased)
35
  * **[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*.
36
 
37
 
38
- ![model image](https://camo.githubusercontent.com/623b4dea0b653f2ad3f36c71ebfe749a677ac0a1/68747470733a2f2f6d69726f2e6d656469756d2e636f6d2f6d61782f343030362f312a44304a31674e51663876727255704b657944387750412e706e67)
39
-
40
-
41
  ## Tokenizer
42
 
43
  The model uses a cased SentencePiece tokenizer configured with the `Nmt, NFKC, Replace multi-space to single-space` normalizers
@@ -45,9 +42,9 @@ and has 32003 tokens.
45
  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).
46
  See [./raw/main/tokenizer.json](tokenizer.json) for details.
47
 
48
- ## Dataset
49
 
50
- All models listed below are trained on
51
  [cleaned Dutch mC4](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned),
52
  which is the original mC4, except
53
 
@@ -58,96 +55,138 @@ which is the original mC4, except
58
  * Documents with "javascript", "lorum ipsum", "terms of use", "privacy policy", "cookie policy", "uses cookies",
59
  "use of cookies", "use cookies", "elementen ontbreken", "deze printversie" are removed.
60
 
61
- The Dutch and English models are trained on a 50/50% mix of Dutch mC4 and English C4.
 
 
62
 
63
- ## Models
64
 
65
- Three types of models have been trained. `t5-base-dutch` is the only model with an original T5 config.
 
66
  The other model types t5-v1.1 and t5-eff have `gated-relu` instead of `relu` as activation function,
67
  and trained with a drop-out of `0.0` unless training would diverge (`t5-v1.1-large-dutch-cased`).
68
- The T5-eff models are models with mostly different numbers of layers. The table will list
69
- the several dimensions of these models. Note that `efficient` is a misnomer for models with few layers,
70
- e.g. `t5-xl-4L-dutch-english-cased`, that is not efficient and one of the worst models on downstream summarization.
71
 
72
- | | 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 |
73
  |:------------------|:----------------|:-----------------------------|:---------------------------|:----------------------------|:-----------------------------------|:----------------------------------------|:-----------------------------|:-------------------------------|:----------------------------------|:-----------------------------------|:--------------------------------------|
74
- | 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 |
75
- | d_model | 768 | 768 | 768 | 1024 | 768 | 768 | 512 | 2048 | 768 | 1024 | 1024 |
76
- | d_ff | 3072 | 2048 | 2048 | 2816 | 2048 | 2048 | 1920 | 5120 | 2560 | 16384 | 4096 |
77
- | num_heads | 12 | 12 | 12 | 16 | 12 | 12 | 8 | 32 | 12 | 32 | 16 |
78
- | d_kv | 64 | 64 | 64 | 64 | 64 | 64 | 64 | 64 | 64 | 128 | 64 |
79
- | num_layers | 12 | 12 | 12 | 24 | 12 | 12 | 24 | 4 | 36 | 8 | 8 |
80
- | num parameters | 223M | 248M | 248M | 783M | 248M | 248M | 250M | 585M | 729M | 1241M | 335M |
81
- | feed_forward_proj | relu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu |
82
- | dropout | 0.1 | 0.0 | 0.0 | 0.1 | 0.0 | 0.0 | 0.0 | 0.1 | 0.0 | 0.0 | 0.0 |
83
- | 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 |
84
- | tr. seq len | 512 | 1024 | 1024 | 512 | 512 | 1024 | 512 | 512 | 512 | 512 | 512 |
85
- | batch size | 128 | 64 | 64 | 64 | 128 | 64 | 128 | 512 | 512 | 64 | 128 |
86
- | total steps | 527500 | 1014525 | 1210154 | 2427498 | 2839630 | 1520k/3397024 | 851852 | 212963 | 212963 | 538k/1703705 | 851850 |
87
- | epochs | 1 | 2 | 2 | 2 | 10 | 4 | 1 | 1 | 1 | 1 | 1 |
88
- | duration | 2d9h | 5d5h | 6d6h | 8d13h | 11d18h | 9d1h | 4d10h | 6d1h | 17d15h | 4d 19h | 3d 23h |
89
- | optimizer | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor |
90
- | lr | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.009 | 0.005 | 0.005 |
91
- | warmup | 10000.0 | 10000.0 | 10000.0 | 10000.0 | 10000.0 | 5000.0 | 20000.0 | 2500.0 | 1000.0 | 1500.0 | 1500.0 |
92
- | eval loss | 1,38 | 1,20 | 0,96 | 1,07 | 1,11 | 1,13 | 1,18 | 1,27 | 1,05 | 1,3019 | 1,15 |
93
- | eval acc | 0,70 | 0,73 | 0,78 | 0,76 | 0,75 | 0,74 | 0,74 | 0,72 | 0,76 | 0,71 | 0,74 |
94
-
95
- ## Evaluation on summarization
96
-
97
- The models below have been evaluated on the summarization downstream task on 50K samples from the CNN Dailymail dataset.
98
- All models were fine-tuned with the AdamW optimizer with a batch size of 128 and constant learning rate of 1e-3 after a
99
- warmup of 64 steps, with a label smoothing factor of 0.05.
100
- Article and summary token lengths were set to 1024 and 142.
101
-
102
- | | 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 |
103
- |:-------------------|:----------------|:-----------------------------|:---------------------------|:-----------------------------------|:----------------------------------------|:-----------------------------|:-------------------------------|:----------------------------------|:--------------------------------------|:-----------|
104
- | rouge1 | 33.0313 | 33.8432 | 34.0906 | 33.1116 | 34.6465 | 34.376 | 30.8983 | 35.0931 | 33.9293 | 33.6466 |
105
- | rouge2 | 12.9452 | 13.7706 | 13.6203 | 13.275 | 13.8525 | 13.8939 | 11.6005 | 14.3823 | 13.6274 | 13.1085 |
106
- | rougeL | 23.7204 | 24.5642 | 24.7304 | 24.3561 | 24.721 | 25.2496 | 22.6536 | 25.3213 | 24.5595 | 23.909 |
107
- | rougeLsum | 29.842 | 30.7783 | 31.1438 | 30.0548 | 31.6104 | 31.3838 | 27.8467 | 32.3526 | 30.952 | 30.5054 |
108
- | gen_len | 90.488 | 91.832 | 92.122 | 89.583 | 98.333 | 90.442 | 92.342 | 96.832 | 95.057 | 96.312 |
109
- | num parameters | 223M | 248M | 248M | 248M | 248M | 250M | 585M | 729M | 335M | 582M |
110
- | samples_per_second | 3.195 | 3.039 | 3.0 | 3.216 | 2.974 | 1.594 | 2.47 | 0.623 | 3.087 | 1.201 |
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
111
 
112
  ## Translation models
113
 
114
- The small 24L and base 36L models have been fine-tuned for translation on the CCMatrix dataset.
115
- The models named *-`multi` support both directions of translation. The models are trained on CCMatrix only. As this is
116
- a really large dataset with over 100M Dutch-English sentence pairs, the models are trained on a fraction of it,
117
- refer to the table below for how long. Evaluation is performed on a CCMatrix section not trained on, but also
118
- on Tatoeba and Opus Books. The `_bp` columns list the *brevity penalty*. The `avg_bleu` score is the bleu score
119
- averaged over all three evaluation datasets.
120
-
121
- The translation metrics are listed in the table below:
122
-
123
- | | 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 |
124
- |:-----------------------|:-----------------------------|:-----------------------------|:-----------------------------|:------------------------------|:------------------------------|
125
- | id | 0 | 14 | 15 | 16 | 20 |
126
- | source_lang | en | en | nl | en | nl |
127
- | target_lang | nl | nl | en | nl | en |
128
- | source_prefix | translate English to Dutch: | translate English to Dutch: | translate Dutch to English: | translate English to Dutch: | translate Dutch to English: |
129
- | tatoeba_bp | 0.9897614370103832 | 0.9736173618072754 | 0.943521164106552 | 0.9760983304454847 | 0.9406676405486575 |
130
- | ccmatrix_bp | 0.9590750786190209 | 0.9536276245543676 | 0.9635673583308255 | 0.9517934939463099 | 0.9585648049711814 |
131
- | opus_books_bp | 0.7478011343203491 | 0.7950194726093107 | 0.9362852511299413 | 0.770498474692027 | 0.8870675076932444 |
132
- | tatoeba_score | 50.63006965176505 | 46.580601850286214 | 52.82030981131822 | 46.419809813946046 | 51.67887417355214 |
133
- | ccmatrix_score | 60.33227938980884 | 56.81297258845844 | 62.836646082246254 | 57.404319674892406 | 63.08633155239932 |
134
- | opus_books_score | 10.405013868050663 | 13.477997378535864 | 24.93113308798125 | 12.927244801365507 | 23.418552148252047 |
135
- | avg_bleu | 40.455787636541515 | 38.95719060576017 | 46.86269632718191 | 38.91712476340132 | 46.0612526247345 |
136
- | total steps | 78125 | 390625 | 390625 | 390625 | 390625 |
137
- | duration | 14h | 101h | 101h | 74h | 74h |
138
- | num_parameters | 728928000 | 728928000 | 728928000 | 249991680 | 249991680 |
139
- | label_smoothing_factor | 0.09 | 0.15 | 0.15 | 0.1 | 0.1 |
140
- | learning_rate | 0.0001 | 5e-05 | 5e-05 | 0.0005 | 0.0005 |
 
 
 
141
 
142
  ## Acknowledgements
143
 
144
  This project would not have been possible without compute generously provided by Google through the
145
- [TPU Research Cloud](https://sites.research.google/trc/). The HuggingFace 🤗 ecosystem and was also
146
- instrumental all parts of the training. Logging metrics to Weights & Biases made it possible to keep track of many
147
- models and orchestrate hyper-parameter sweeps with insightful visualizations. I cannot imagine how I would
148
- have completed this project otherwise.
149
  The following repositories where helpful in setting up the TPU-VM,
150
- and getting an idea what sensible hyper-parameters are for training gpt2 from scratch.
151
 
152
  * [Gsarti's Pretrain and Fine-tune a T5 model with Flax on GCP](https://github.com/gsarti/t5-flax-gcp)
153
  * [Flax/Jax Community week t5-base-dutch](https://huggingface.co/flax-community/t5-base-dutch)
 
17
  pre-trained from scratch on [cleaned Dutch 🇳🇱🇧🇪 mC4](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned).
18
 
19
 
20
+
21
+
22
  This **t5-v1.1** model has **247M** parameters.
23
+ It was pre-trained with masked language modeling (denoise token span corruption) objective on the dataset
24
  `mc4_nl_cleaned` config `full` for **2** epoch(s) and a duration of **6d6h**,
25
+ with a sequence length of **1024**, batch size **64** and **1210154** total steps (**79B** tokens).
26
  Pre-training evaluation loss and accuracy are **0,96** and **0,78**.
27
+ Refer to the evaluation section below for a comparison of the pre-trained models on summarization and translation.
 
 
28
  * 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.
29
  * For a demo of the Dutch CNN summarization models, head over to the Hugging Face Spaces for
30
  the **[Netherformer 📰](https://huggingface.co/spaces/flax-community/netherformer)** example application!
 
35
  * **[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*.
36
 
37
 
 
 
 
38
  ## Tokenizer
39
 
40
  The model uses a cased SentencePiece tokenizer configured with the `Nmt, NFKC, Replace multi-space to single-space` normalizers
 
42
  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).
43
  See [./raw/main/tokenizer.json](tokenizer.json) for details.
44
 
45
+ ## Dataset(s)
46
 
47
+ All models listed below are pre-trained on
48
  [cleaned Dutch mC4](https://huggingface.co/datasets/yhavinga/mc4_nl_cleaned),
49
  which is the original mC4, except
50
 
 
55
  * Documents with "javascript", "lorum ipsum", "terms of use", "privacy policy", "cookie policy", "uses cookies",
56
  "use of cookies", "use cookies", "elementen ontbreken", "deze printversie" are removed.
57
 
58
+ The Dutch and English models are pre-trained on a 50/50% mix of Dutch mC4 and English C4.
59
+
60
+ The translation models are fine-tuned on [CCMatrix](https://huggingface.co/datasets/yhavinga/ccmatrix).
61
 
62
+ ## Dutch T5 Models
63
 
64
+ Three types of [Dutch T5 models have been trained (blog)](https://huggingface.co/spaces/yhavinga/pre-training-dutch-t5-models).
65
+ `t5-base-dutch` is the only model with an original T5 config.
66
  The other model types t5-v1.1 and t5-eff have `gated-relu` instead of `relu` as activation function,
67
  and trained with a drop-out of `0.0` unless training would diverge (`t5-v1.1-large-dutch-cased`).
68
+ The T5-eff models are models that differ in their number of layers. The table will list
69
+ the several dimensions of these models. Not all t5-eff models are efficient, the best example being the inefficient
70
+ `t5-xl-4L-dutch-english-cased`.
71
 
72
+ | | [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) |
73
  |:------------------|:----------------|:-----------------------------|:---------------------------|:----------------------------|:-----------------------------------|:----------------------------------------|:-----------------------------|:-------------------------------|:----------------------------------|:-----------------------------------|:--------------------------------------|
74
+ | *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 |
75
+ | *d_model* | 768 | 768 | 768 | 1024 | 768 | 768 | 512 | 2048 | 768 | 1024 | 1024 |
76
+ | *d_ff* | 3072 | 2048 | 2048 | 2816 | 2048 | 2048 | 1920 | 5120 | 2560 | 16384 | 4096 |
77
+ | *num_heads* | 12 | 12 | 12 | 16 | 12 | 12 | 8 | 32 | 12 | 32 | 16 |
78
+ | *d_kv* | 64 | 64 | 64 | 64 | 64 | 64 | 64 | 64 | 64 | 128 | 64 |
79
+ | *num_layers* | 12 | 12 | 12 | 24 | 12 | 12 | 24 | 4 | 36 | 8 | 8 |
80
+ | *num parameters* | 223M | 248M | 248M | 783M | 248M | 248M | 250M | 585M | 729M | 1241M | 335M |
81
+ | *feed_forward_proj* | relu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu | gated-gelu |
82
+ | *dropout* | 0.1 | 0.0 | 0.0 | 0.1 | 0.0 | 0.0 | 0.0 | 0.1 | 0.0 | 0.0 | 0.0 |
83
+ | *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 |
84
+ | *tr. seq len* | 512 | 1024 | 1024 | 512 | 512 | 1024 | 512 | 512 | 512 | 512 | 512 |
85
+ | *batch size* | 128 | 64 | 64 | 64 | 128 | 64 | 128 | 512 | 512 | 64 | 128 |
86
+ | *total steps* | 527500 | 1014525 | 1210154 | 1120k/2427498 | 2839630 | 1520k/3397024 | 851852 | 212963 | 212963 | 538k/1703705 | 851850 |
87
+ | *epochs* | 1 | 2 | 2 | 2 | 10 | 4 | 1 | 1 | 1 | 1 | 1 |
88
+ | *duration* | 2d9h | 5d5h | 6d6h | 8d13h | 11d18h | 9d1h | 4d10h | 6d1h | 17d15h | 4d 19h | 3d 23h |
89
+ | *optimizer* | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor | adafactor |
90
+ | *lr* | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.005 | 0.009 | 0.005 | 0.005 |
91
+ | *warmup* | 10000.0 | 10000.0 | 10000.0 | 10000.0 | 10000.0 | 5000.0 | 20000.0 | 2500.0 | 1000.0 | 1500.0 | 1500.0 |
92
+ | *eval loss* | 1,38 | 1,20 | 0,96 | 1,07 | 1,11 | 1,13 | 1,18 | 1,27 | 1,05 | 1,3019 | 1,15 |
93
+ | *eval acc* | 0,70 | 0,73 | 0,78 | 0,76 | 0,75 | 0,74 | 0,74 | 0,72 | 0,76 | 0,71 | 0,74 |
94
+
95
+ ## Evaluation
96
+
97
+ Most models from the list above have been fine-tuned for summarization and translation.
98
+ The figure below shows the evaluation scores, where the x-axis shows the translation Bleu score (higher is better)
99
+ and y-axis the summarization Rouge1 translation score (higher is better).
100
+ Point size is proportional to the model size. Models with faster inference speed are green, slower inference speed is
101
+ plotted as bleu.
102
+
103
+ ![Evaluation T5 Dutch English](evaluation_t5_dutch_english.png)
104
+
105
+ Evaluation was run on fine-tuned models trained with the following settings:
106
+
107
+
108
+ | | Summarization | Translation |
109
+ |---------------:|------------------|-------------------|
110
+ | Dataset | CNN Dailymail NL | CCMatrix en -> nl |
111
+ | #train samples | 50K | 50K |
112
+ | Optimizer | Adam | Adam |
113
+ | learning rate | 0.001 | 0.0005 |
114
+ | source length | 1024 | 128 |
115
+ | target length | 142 | 128 |
116
+ |label smoothing | 0.05 | 0.1 |
117
+ | #eval samples | 1000 | 1000 |
118
+
119
+ Note that the amount of training data is limited to a fraction of the total dataset sizes, therefore the scores
120
+ below can only be used to compare the 'transfer-learning' strength. The fine-tuned checkpoints for this evaluation
121
+ are not saved, since they were trained for comparison of pre-trained models only.
122
+
123
+ The numbers for summarization are the Rouge scores on 1000 documents from the test split.
124
+
125
+ | | [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 |
126
+ |:------------------------|----------------:|-----------------------------:|---------------------------:|-----------------------------------:|----------------------------------------:|-----------------------------:|-------------------------------:|----------------------------------:|--------------------------------------:|-----------:|
127
+ | *rouge1* | 33.38 | 33.97 | 34.39 | 33.38 | 34.97 | 34.38 | 30.35 | **35.04** | 34.04 | 33.25 |
128
+ | *rouge2* | 13.32 | 13.85 | 13.98 | 13.47 | 14.01 | 13.89 | 11.57 | **14.23** | 13.76 | 12.74 |
129
+ | *rougeL* | 24.22 | 24.72 | 25.1 | 24.34 | 24.99 | **25.25** | 22.69 | 25.05 | 24.75 | 23.5 |
130
+ | *rougeLsum* | 30.23 | 30.9 | 31.44 | 30.51 | 32.01 | 31.38 | 27.5 | **32.12** | 31.12 | 30.15 |
131
+ | *samples_per_second* | 3.18 | 3.02 | 2.99 | 3.22 | 2.97 | 1.57 | 2.8 | 0.61 | **3.27** | 1.22 |
132
+
133
+ The models below have been evaluated for English to Dutch translation.
134
+ Note that the first four models are pre-trained on Dutch only. That they still perform adequate is probably because
135
+ the translation direction is English to Dutch.
136
+ The numbers reported are the Bleu scores on 1000 documents from the test split.
137
+
138
+ | | [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 |
139
+ |:-------------------------------|----------------:|-----------------------------:|---------------------------:|----------------------------:|-----------------------------------:|----------------------------------------:|-----------------------------:|-------------------------------:|----------------------------------:|--------------------------------------:|-----------:|
140
+ | *precision_ng1* | 74.17 | 78.09 | 77.08 | 72.12 | 77.19 | 78.76 | 78.59 | 77.3 | **79.75** | 78.88 | 73.47 |
141
+ | *precision_ng2* | 52.42 | 57.52 | 55.31 | 48.7 | 55.39 | 58.01 | 57.83 | 55.27 | **59.89** | 58.27 | 50.12 |
142
+ | *precision_ng3* | 39.55 | 45.2 | 42.54 | 35.54 | 42.25 | 45.13 | 45.02 | 42.06 | **47.4** | 45.95 | 36.59 |
143
+ | *precision_ng4* | 30.23 | 36.04 | 33.26 | 26.27 | 32.74 | 35.72 | 35.41 | 32.61 | **38.1** | 36.91 | 27.26 |
144
+ | *bp* | 0.99 | 0.98 | 0.97 | 0.98 | 0.98 | 0.98 | 0.98 | 0.97 | 0.98 | 0.98 | 0.98 |
145
+ | *score* | 45.88 | 51.21 | 48.31 | 41.59 | 48.17 | 51.31 | 50.82 | 47.83 | **53** | 51.79 | 42.74 |
146
+ | *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 |
147
+
148
 
149
  ## Translation models
150
 
151
+ The models `t5-small-24L-dutch-english` and `t5-base-36L-dutch-english` have been fine-tuned for both language
152
+ directions on the first 25M samples from CCMatrix, giving a total of 50M training samples.
153
+ Evaluation is performed on out-of-sample CCMatrix and also on Tatoeba and Opus Books.
154
+ The `_bp` columns list the *brevity penalty*. The `avg_bleu` score is the bleu score
155
+ averaged over all three evaluation datasets. The best scores displayed in bold for both translation directions.
156
+
157
+ | | [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) |
158
+ |:-----------------------|:-----------------------------|:-----------------------------|:------------------------------|:------------------------------|
159
+ | *source_lang* | en | nl | en | nl |
160
+ | *target_lang* | nl | en | nl | en |
161
+ | *source_prefix* | translate English to Dutch: | translate Dutch to English: | translate English to Dutch: | translate Dutch to English: |
162
+ | *ccmatrix_bleu* | **56.8** | 62.8 | 57.4 | **63.1** |
163
+ | *tatoeba_bleu* | **46.6** | **52.8** | 46.4 | 51.7 |
164
+ | *opus_books_bleu* | **13.5** | **24.9** | 12.9 | 23.4 |
165
+ | *ccmatrix_bp* | 0.95 | 0.96 | 0.95 | 0.96 |
166
+ | *tatoeba_bp* | 0.97 | 0.94 | 0.98 | 0.94 |
167
+ | *opus_books_bp* | 0.8 | 0.94 | 0.77 | 0.89 |
168
+ | *avg_bleu* | **38.96** | **46.86** | 38.92 | 46.06 |
169
+ | *max_source_length* | 128 | 128 | 128 | 128 |
170
+ | *max_target_length* | 128 | 128 | 128 | 128 |
171
+ | *adam_beta1* | 0.9 | 0.9 | 0.9 | 0.9 |
172
+ | *adam_beta2* | 0.997 | 0.997 | 0.997 | 0.997 |
173
+ | *weight_decay* | 0.05 | 0.05 | 0.002 | 0.002 |
174
+ | *lr* | 5e-05 | 5e-05 | 0.0005 | 0.0005 |
175
+ | *label_smoothing_factor* | 0.15 | 0.15 | 0.1 | 0.1 |
176
+ | *train_batch_size* | 128 | 128 | 128 | 128 |
177
+ | *warmup_steps* | 2000 | 2000 | 2000 | 2000 |
178
+ | *total steps* | 390625 | 390625 | 390625 | 390625 |
179
+ | *duration* | 4d 5h | 4d 5h | 3d 2h | 3d 2h |
180
+ | *num parameters* | 729M | 729M | 250M | 250M |
181
 
182
  ## Acknowledgements
183
 
184
  This project would not have been possible without compute generously provided by Google through the
185
+ [TPU Research Cloud](https://sites.research.google/trc/). The HuggingFace 🤗 ecosystem was instrumental in all parts
186
+ of the training. Weights & Biases made it possible to keep track of many training sessions
187
+ and orchestrate hyper-parameter sweeps with insightful visualizations.
 
188
  The following repositories where helpful in setting up the TPU-VM,
189
+ and getting an idea what sensible hyper-parameters are for training gpt2 from scratch:
190
 
191
  * [Gsarti's Pretrain and Fine-tune a T5 model with Flax on GCP](https://github.com/gsarti/t5-flax-gcp)
192
  * [Flax/Jax Community week t5-base-dutch](https://huggingface.co/flax-community/t5-base-dutch)
evaluation_t5_dutch_english.png ADDED