stefan-it commited on
Commit
75a7c2b
1 Parent(s): f29e458

readme: add initial version (#1)

Browse files

- readme: add initial version (589599dd22022126e53adcb3dd057edf7c698d3e)

Files changed (1) hide show
  1. README.md +84 -0
README.md ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: fr
3
+ license: mit
4
+ tags:
5
+ - flair
6
+ - token-classification
7
+ - sequence-tagger-model
8
+ base_model: hmbyt5-preliminary/byt5-small-historic-multilingual-span20-flax
9
+ inference: false
10
+ widget:
11
+ - text: 'Parmi les remèdes recommandés par la Société , il faut mentionner celui que
12
+ M . Schatzmann , de Lausanne , a proposé :'
13
+ ---
14
+
15
+ # Fine-tuned Flair Model on LeTemps French NER Dataset (HIPE-2022)
16
+
17
+ This Flair model was fine-tuned on the
18
+ [LeTemps French](https://github.com/hipe-eval/HIPE-2022-data/blob/main/documentation/README-letemps.md)
19
+ NER Dataset using hmByT5 as backbone LM.
20
+
21
+ The LeTemps dataset consists of NE-annotated historical French newspaper articles from mid-19C to mid 20C.
22
+
23
+ The following NEs were annotated: `loc`, `org` and `pers`.
24
+
25
+ # ⚠️ Inference Widget ⚠️
26
+
27
+ Fine-Tuning ByT5 models in Flair is currently done by implementing an own [`ByT5Embedding`][1] class.
28
+
29
+ This class needs to be present when running the model with Flair.
30
+
31
+ Thus, the inference widget is not working with hmByT5 at the moment on the Model Hub and is currently disabled.
32
+
33
+ This should be fixed in future, when ByT5 fine-tuning is supported in Flair directly.
34
+
35
+ [1]: https://github.com/stefan-it/hmBench/blob/main/byt5_embeddings.py
36
+
37
+ # Results
38
+
39
+ We performed a hyper-parameter search over the following parameters with 5 different seeds per configuration:
40
+
41
+ * Batch Sizes: `[8, 4]`
42
+ * Learning Rates: `[0.00015, 0.00016]`
43
+
44
+ And report micro F1-score on development set:
45
+
46
+ | Configuration | Run 1 | Run 2 | Run 3 | Run 4 | Run 5 | Avg. |
47
+ |-------------------|--------------|--------------|--------------|--------------|--------------|--------------|
48
+ | bs8-e10-lr0.00016 | [0.6553][1] | [0.6628][2] | [0.6699][3] | [0.6524][4] | [0.6542][5] | 65.89 ± 0.65 |
49
+ | bs4-e10-lr0.00015 | [0.6603][6] | [0.6651][7] | [0.654][8] | [0.6575][9] | [0.6575][10] | 65.89 ± 0.37 |
50
+ | bs4-e10-lr0.00016 | [0.6423][11] | [0.6595][12] | [0.6625][13] | [0.6657][14] | [0.6538][15] | 65.68 ± 0.82 |
51
+ | bs8-e10-lr0.00015 | [0.6502][16] | [0.6541][17] | [0.6607][18] | [0.6496][19] | [0.6629][20] | 65.55 ± 0.54 |
52
+
53
+ [1]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-1
54
+ [2]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-2
55
+ [3]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-3
56
+ [4]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-4
57
+ [5]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-5
58
+ [6]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-1
59
+ [7]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-2
60
+ [8]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-3
61
+ [9]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-4
62
+ [10]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-5
63
+ [11]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-1
64
+ [12]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-2
65
+ [13]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-3
66
+ [14]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-4
67
+ [15]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-5
68
+ [16]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-1
69
+ [17]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-2
70
+ [18]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-3
71
+ [19]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-4
72
+ [20]: https://hf.co/hmbench/hmbench-letemps-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-5
73
+
74
+ The [training log](training.log) and TensorBoard logs are also uploaded to the model hub.
75
+
76
+ More information about fine-tuning can be found [here](https://github.com/stefan-it/hmBench).
77
+
78
+ # Acknowledgements
79
+
80
+ We thank [Luisa März](https://github.com/LuisaMaerz), [Katharina Schmid](https://github.com/schmika) and
81
+ [Erion Çano](https://github.com/erionc) for their fruitful discussions about Historic Language Models.
82
+
83
+ Research supported with Cloud TPUs from Google's [TPU Research Cloud](https://sites.research.google/trc/about/) (TRC).
84
+ Many Thanks for providing access to the TPUs ❤️