Hugging Face's logo

stefan-it
/
hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-1

Token Classification
Flair
PyTorch
TensorBoard
Dutch
sequence-tagger-model
Model card Files Metrics Community
stefan-it commited on
Commit
2ee81e5
•
1 Parent(s): 1f0fede

readme: add initial version of model card

Browse files

Hey,

this commit adds the initial version of model card.

Files changed (1) hide show
  1. README.md +74 -0
README.md ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: nl
3
+ license: mit
4
+ tags:
5
+ - flair
6
+ - token-classification
7
+ - sequence-tagger-model
8
+ base_model: dbmdz/bert-tiny-historic-multilingual-cased
9
+ widget:
10
+ - text: Professoren der Geneeskun dige Faculteit te Groningen alsook van de HH , Doctoren
11
+ en Chirurgijns van Groningen , Friesland , Noordholland , Overijssel , Gelderland
12
+ , Drenthe , in welke Provinciën dit Elixir als Medicament voor Mond en Tanden
13
+ reeds jaren bakend is .
14
+ ---
15
+
16
+ # Fine-tuned Flair Model on Dutch ICDAR-Europeana NER Dataset
17
+
18
+ This Flair model was fine-tuned on the
19
+ [Dutch ICDAR-Europeana](https://github.com/stefan-it/historic-domain-adaptation-icdar)
20
+ NER Dataset using hmBERT Tiny as backbone LM.
21
+
22
+ The ICDAR-Europeana NER Dataset is a preprocessed variant of the
23
+ [Europeana NER Corpora](https://github.com/EuropeanaNewspapers/ner-corpora) for Dutch and French.
24
+
25
+ The following NEs were annotated: `PER`, `LOC` and `ORG`.
26
+
27
+ # Results
28
+
29
+ We performed a hyper-parameter search over the following parameters with 5 different seeds per configuration:
30
+
31
+ * Batch Sizes: `[4, 8]`
32
+ * Learning Rates: `[5e-05, 3e-05]`
33
+
34
+ And report micro F1-score on development set:
35
+
36
+ | Configuration | Seed 1 | Seed 2 | Seed 3 | Seed 4 | Seed 5 | Average |
37
+ |-------------------|-----------------|--------------|--------------|--------------|--------------|-----------------|
38
+ | `bs4-e10-lr5e-05` | [0.5572][1] | [0.5434][2] | [0.5984][3] | [0.5636][4] | [0.5674][5] | 0.566 ± 0.0203 |
39
+ | `bs8-e10-lr5e-05` | [**0.5072**][6] | [0.5287][7] | [0.5641][8] | [0.5438][9] | [0.5346][10] | 0.5357 ± 0.0208 |
40
+ | `bs4-e10-lr3e-05` | [0.519][11] | [0.471][12] | [0.5479][13] | [0.498][14] | [0.4977][15] | 0.5067 ± 0.0286 |
41
+ | `bs8-e10-lr3e-05` | [0.4817][16] | [0.4511][17] | [0.4956][18] | [0.4627][19] | [0.4715][20] | 0.4725 ± 0.0171 |
42
+
43
+ [1]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs4-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-1
44
+ [2]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs4-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-2
45
+ [3]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs4-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-3
46
+ [4]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs4-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-4
47
+ [5]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs4-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-5
48
+ [6]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-1
49
+ [7]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-2
50
+ [8]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-3
51
+ [9]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-4
52
+ [10]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr5e-05-poolingfirst-layers-1-crfFalse-5
53
+ [11]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs4-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-1
54
+ [12]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs4-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-2
55
+ [13]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs4-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-3
56
+ [14]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs4-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-4
57
+ [15]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs4-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-5
58
+ [16]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-1
59
+ [17]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-2
60
+ [18]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-3
61
+ [19]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-4
62
+ [20]: https://hf.co/stefan-it/hmbench-icdar-nl-hmbert_tiny-bs8-wsFalse-e10-lr3e-05-poolingfirst-layers-1-crfFalse-5
63
+
64
+ The [training log](training.log) and TensorBoard logs (not available for hmBERT Base model) are also uploaded to the model hub.
65
+
66
+ More information about fine-tuning can be found [here](https://github.com/stefan-it/hmBench).
67
+
68
+ # Acknowledgements
69
+
70
+ We thank [Luisa März](https://github.com/LuisaMaerz), [Katharina Schmid](https://github.com/schmika) and
71
+ [Erion Çano](https://github.com/erionc) for their fruitful discussions about Historic Language Models.
72
+
73
+ Research supported with Cloud TPUs from Google's [TPU Research Cloud](https://sites.research.google/trc/about/) (TRC).
74
+ Many Thanks for providing access to the TPUs ❤️