|
--- |
|
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 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/) |
|
|
|
|
