t5-base-dutch
Created by Yeb Havinga & Dat Nguyen during the Hugging Face community week, organized by HuggingFace and TPU usage sponsored by Google, for the project Pre-train T5 from scratch in Dutch. See also the fine-tuned t5-base-dutch-demo model, and the demo application Netherformer π°, 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 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 π° 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 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 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.
See ./raw/main/tokenizer.json for details.
Dataset(s)
All models listed below are pre-trained on cleaned Dutch mC4, 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 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.
Dutch T5 Models
Three types of Dutch T5 models have been trained (blog).
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 | 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 | |
---|---|---|---|---|---|---|---|---|---|---|---|
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 | 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 | |
---|---|---|---|---|---|---|---|---|---|---|
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 | 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-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 | t5-base-36L-ccmatrix-multi | t5-small-24L-ccmatrix-multi | 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. The HuggingFace π€ 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:
Created by Yeb Havinga
- Downloads last month
- 64