readme: add initial version

#1
by stefan-it - opened
Files changed (1) hide show
  1. README.md +87 -0
README.md ADDED
@@ -0,0 +1,87 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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: Le Moniteur universel fait ressortir les avantages de la situation de l '
12
+ Allemagne , sa force militaire , le peu d ' intérêts personnels qu ' elle peut
13
+ avoir dans la question d ' Orient .
14
+ ---
15
+
16
+ # Fine-tuned Flair Model on French NewsEye NER Dataset (HIPE-2022)
17
+
18
+ This Flair model was fine-tuned on the
19
+ [French NewsEye](https://github.com/hipe-eval/HIPE-2022-data/blob/main/documentation/README-newseye.md)
20
+ NER Dataset using hmByT5 as backbone LM.
21
+
22
+ The NewsEye dataset is comprised of diachronic historical newspaper material published between 1850 and 1950
23
+ in French, German, Finnish, and Swedish.
24
+ More information can be found [here](https://dl.acm.org/doi/abs/10.1145/3404835.3463255).
25
+
26
+ The following NEs were annotated: `PER`, `LOC`, `ORG` and `HumanProd`.
27
+
28
+ # ⚠️ Inference Widget ⚠️
29
+
30
+ Fine-Tuning ByT5 models in Flair is currently done by implementing an own [`ByT5Embedding`][1] class.
31
+
32
+ This class needs to be present when running the model with Flair.
33
+
34
+ Thus, the inference widget is not working with hmByT5 at the moment on the Model Hub and is currently disabled.
35
+
36
+ This should be fixed in future, when ByT5 fine-tuning is supported in Flair directly.
37
+
38
+ [1]: https://github.com/stefan-it/hmBench/blob/main/byt5_embeddings.py
39
+
40
+ # Results
41
+
42
+ We performed a hyper-parameter search over the following parameters with 5 different seeds per configuration:
43
+
44
+ * Batch Sizes: `[8, 4]`
45
+ * Learning Rates: `[0.00015, 0.00016]`
46
+
47
+ And report micro F1-score on development set:
48
+
49
+ | Configuration | Run 1 | Run 2 | Run 3 | Run 4 | Run 5 | Avg. |
50
+ |-------------------|--------------|--------------|--------------|--------------|--------------|--------------|
51
+ | bs4-e10-lr0.00016 | [0.793][1] | [0.803][2] | [0.8054][3] | [0.8069][4] | [0.8133][5] | 80.43 ± 0.66 |
52
+ | bs8-e10-lr0.00016 | [0.7888][6] | [0.8094][7] | [0.8043][8] | [0.8011][9] | [0.8117][10] | 80.31 ± 0.8 |
53
+ | bs4-e10-lr0.00015 | [0.7884][11] | [0.8109][12] | [0.8005][13] | [0.8083][14] | [0.8022][15] | 80.21 ± 0.78 |
54
+ | bs8-e10-lr0.00015 | [0.788][16] | [0.8003][17] | [0.8067][18] | [0.8035][19] | [0.8064][20] | 80.1 ± 0.69 |
55
+
56
+ [1]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-1
57
+ [2]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-2
58
+ [3]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-3
59
+ [4]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-4
60
+ [5]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-5
61
+ [6]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-1
62
+ [7]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-2
63
+ [8]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-3
64
+ [9]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-4
65
+ [10]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00016-poolingfirst-layers-1-crfFalse-5
66
+ [11]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-1
67
+ [12]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-2
68
+ [13]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-3
69
+ [14]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-4
70
+ [15]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs4-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-5
71
+ [16]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-1
72
+ [17]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-2
73
+ [18]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-3
74
+ [19]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-4
75
+ [20]: https://hf.co/hmbench/hmbench-newseye-fr-hmbyt5-bs8-wsFalse-e10-lr0.00015-poolingfirst-layers-1-crfFalse-5
76
+
77
+ The [training log](training.log) and TensorBoard logs are also uploaded to the model hub.
78
+
79
+ More information about fine-tuning can be found [here](https://github.com/stefan-it/hmBench).
80
+
81
+ # Acknowledgements
82
+
83
+ We thank [Luisa März](https://github.com/LuisaMaerz), [Katharina Schmid](https://github.com/schmika) and
84
+ [Erion Çano](https://github.com/erionc) for their fruitful discussions about Historic Language Models.
85
+
86
+ Research supported with Cloud TPUs from Google's [TPU Research Cloud](https://sites.research.google/trc/about/) (TRC).
87
+ Many Thanks for providing access to the TPUs ❤️