readme: add initial version
Browse files
README.md
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
- de
|
| 6 |
+
- fr
|
| 7 |
+
- fi
|
| 8 |
+
- sv
|
| 9 |
+
- nl
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# hmByT5 - Preliminary Language Models
|
| 13 |
+
|
| 14 |
+
Preliminary Historic Multilingual and Monolingual ByT5 Models. Following languages are currently covered:
|
| 15 |
+
|
| 16 |
+
* English (British Library Corpus - Books)
|
| 17 |
+
* German (Europeana Newspaper)
|
| 18 |
+
* French (Europeana Newspaper)
|
| 19 |
+
* Finnish (Europeana Newspaper)
|
| 20 |
+
* Swedish (Europeana Newspaper)
|
| 21 |
+
* Dutch (Delpher Corpus)
|
| 22 |
+
|
| 23 |
+
More details can be found in [our GitHub repository](https://github.com/stefan-it/hmByT5).
|
| 24 |
+
|
| 25 |
+
In this experiment we sample 4B bytes (~4GB of text) from each corpora (and upsample Swedish and Finnish) and train for another epoch (3 epochs in total).
|
| 26 |
+
|
| 27 |
+
# Pretraining
|
| 28 |
+
|
| 29 |
+
We use the official JAX/FLAX example in Hugging Face Transformers to pretrain a ByT5 model on a single v3-8 TPU.
|
| 30 |
+
Details about the training can be found [here](https://github.com/stefan-it/hmByT5/tree/main/hmbyt5-flax).
|
| 31 |
+
|
| 32 |
+
# Evaluation on Downstream Tasks (NER)
|
| 33 |
+
|
| 34 |
+
We evaluated the hmByT5 model on downstream tasks:
|
| 35 |
+
|
| 36 |
+
| Model | English AjMC | German AjMC | French AjMC | Finnish NewsEye | Swedish NewsEye | Dutch ICDAR | French ICDAR | Avg. |
|
| 37 |
+
|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------|--------------|--------------|-----------------|-----------------|--------------|--------------|------|
|
| 38 |
+
| [`hmbyt5-preliminary/byt5-small-multilingual-4g-3e`](https://huggingface.co/hmbyt5-preliminary/byt5-small-multilingual-4g-3e) | 83.49 ± 0.99 | 87.38 ± 0.53 | 84.30 ± 0.51 | | | | | |
|
| 39 |
+
|
| 40 |
+
# Acknowledgements
|
| 41 |
+
|
| 42 |
+
Research supported with Cloud TPUs from Google's [TPU Research Cloud](https://sites.research.google/trc/about/) (TRC).
|
| 43 |
+
Many Thanks for providing access to the TPUs ❤️
|