t5-v1.1-base-dutch-cased

A T5 sequence to sequence model pre-trained from scratch on cleaned Dutch 🇳🇱🇧🇪 mC4.

This t5-v1.1 model has 247M parameters. It was pre-trained with masked language modeling (denoise token span corruption) objective on the dataset mc4_nl_cleaned config full for 2 epoch(s) and a duration of 6d6h, with a sequence length of 1024, batch size 64 and 1210154 total steps (79B tokens). Pre-training evaluation loss and accuracy are 0,96 and 0,78. 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 📰 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-v1.1-base-dutch-cased) 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 🤗 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

yhavinga
/

t5-v1.1-base-dutch-cased