t5 seq2seq

t5-eff-xl-8l-dutch-english-cased

A T5 sequence to sequence model pre-trained from scratch on cleaned Dutch 🇳🇱🇧🇪 mC4 and cleaned English 🇬🇧 C4.

This t5 eff model has 1240M parameters. It was pre-trained with masked language modeling (denoise token span corruption) objective on the dataset mc4_nl_cleaned config large_en_nl for 1 epoch(s) and a duration of 4d 19h, with a sequence length of 512, batch size 64 and 538k/1703705 total steps (18B tokens). Pre-training evaluation loss and accuracy are 1,3019 and 0,71.

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-eff-xl-8l-dutch-english-cased) is unrelated to these projects and not an 'official' checkpoint.

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 and English 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

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 T5 Dutch English

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