stefan-it's picture
readme: fix link reference for ByT5 embedding implementation
d75f975
metadata
language: fr
license: mit
tags:
  - flair
  - token-classification
  - sequence-tagger-model
base_model: hmbyt5-preliminary/byt5-small-historic-multilingual-span20-flax
inference: false
widget:
  - text: >-
      Je suis convaincu , a-t43 dit . que nous n"y parviendrions pas , mais nous
      ne pouvons céder parce que l' état moral de nos troupe* en souffrirait
      trop . ( Fournier . ) Des avions ennemis lancent dix-sept bombes sur
      Dunkerque LONDRES . 31 décembre .

Fine-tuned Flair Model on French ICDAR-Europeana NER Dataset

This Flair model was fine-tuned on the French ICDAR-Europeana NER Dataset using hmByT5 as backbone LM.

The ICDAR-Europeana NER Dataset is a preprocessed variant of the Europeana NER Corpora for Dutch and French.

The following NEs were annotated: PER, LOC and ORG.

⚠️ Inference Widget ⚠️

Fine-Tuning ByT5 models in Flair is currently done by implementing an own ByT5Embedding class.

This class needs to be present when running the model with Flair.

Thus, the inference widget is not working with hmByT5 at the moment on the Model Hub and is currently disabled.

This should be fixed in future, when ByT5 fine-tuning is supported in Flair directly.

Results

We performed a hyper-parameter search over the following parameters with 5 different seeds per configuration:

  • Batch Sizes: [8, 4]
  • Learning Rates: [0.00015, 0.00016]

And report micro F1-score on development set:

Configuration Run 1 Run 2 Run 3 Run 4 Run 5 Avg.
bs8-e10-lr0.00015 0.7743 0.7654 0.7604 0.7729 0.7736 76.93 ± 0.55
bs8-e10-lr0.00016 0.7686 0.7648 0.7678 0.7653 0.7755 76.84 ± 0.38
bs4-e10-lr0.00015 0.7757 0.7549 0.7693 0.7597 0.7696 76.58 ± 0.75
bs4-e10-lr0.00016 0.7625 0.7575 0.769 0.7635 0.7647 76.34 ± 0.37

The training log and TensorBoard logs (only for hmByT5 and hmTEAMS based models) are also uploaded to the model hub.

More information about fine-tuning can be found here.

Acknowledgements

We thank Luisa März, Katharina Schmid and Erion Çano for their fruitful discussions about Historic Language Models.

Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC). Many Thanks for providing access to the TPUs ❤️