readme: add initial version of model card

#1
Files changed (1) hide show
  1. README.md +84 -0
README.md ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: de
3
+ license: mit
4
+ tags:
5
+ - flair
6
+ - token-classification
7
+ - sequence-tagger-model
8
+ - hetzner
9
+ - hetzner-gex44
10
+ - hetzner-gpu
11
+ base_model: dbmdz/bert-base-german-cased
12
+ widget:
13
+ - text: Wesentliche Tätigkeiten der Compliance-Funktion wurden an die Mercurtainment
14
+ AG , Düsseldorf , ausgelagert .
15
+ ---
16
+
17
+ # Fine-tuned Flair Model on CO-Fun NER Dataset
18
+
19
+ This Flair model was fine-tuned on the
20
+ [CO-Fun](https://arxiv.org/abs/2403.15322) NER Dataset using German DBMDZ BERT as backbone LM.
21
+
22
+ ## Dataset
23
+
24
+ The [Company Outsourcing in Fund Prospectuses (CO-Fun) dataset](https://arxiv.org/abs/2403.15322) consists of
25
+ 948 sentences with 5,969 named entity annotations, including 2,340 Outsourced Services, 2,024 Companies, 1,594 Locations
26
+ and 11 Software annotations.
27
+
28
+ Overall, the following named entities are annotated:
29
+
30
+ * `Auslagerung` (engl. outsourcing)
31
+ * `Unternehmen` (engl. company)
32
+ * `Ort` (engl. location)
33
+ * `Software`
34
+
35
+ ## Fine-Tuning
36
+
37
+ The latest [Flair version](https://github.com/flairNLP/flair/tree/42ea3f6854eba04387c38045f160c18bdaac07dc) is used for
38
+ fine-tuning.
39
+
40
+ A hyper-parameter search over the following parameters with 5 different seeds per configuration is performed:
41
+
42
+ * Batch Sizes: [`16`, `8`]
43
+ * Learning Rates: [`3e-05`, `5e-05`]
44
+
45
+ More details can be found in this [repository](https://github.com/stefan-it/co-funer). All models are fine-tuned on a
46
+ [Hetzner GX44](https://www.hetzner.com/dedicated-rootserver/matrix-gpu/) with an NVIDIA RTX 4000.
47
+
48
+ ## Results
49
+
50
+ A hyper-parameter search with 5 different seeds per configuration is performed and micro F1-score on development set
51
+ is reported:
52
+
53
+ | Configuration | Seed 1 | Seed 2 | Seed 3 | Seed 4 | Seed 5 | Average |
54
+ |--------------------|--------------|----------------|--------------|--------------|--------------|-----------------|
55
+ | `bs8-e10-lr5e-05` | [0.9378][1] | [**0.928**][2] | [0.9383][3] | [0.9374][4] | [0.9364][5] | 0.9356 ± 0.0043 |
56
+ | `bs8-e10-lr3e-05` | [0.9336][6] | [0.9366][7] | [0.9299][8] | [0.9417][9] | [0.9281][10] | 0.934 ± 0.0054 |
57
+ | `bs16-e10-lr5e-05` | [0.927][11] | [0.9341][12] | [0.9372][13] | [0.9283][14] | [0.9329][15] | 0.9319 ± 0.0042 |
58
+ | `bs16-e10-lr3e-05` | [0.9141][16] | [0.9321][17] | [0.9175][18] | [0.9391][19] | [0.9177][20] | 0.9241 ± 0.0109 |
59
+
60
+ [1]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr5e-05-1
61
+ [2]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr5e-05-2
62
+ [3]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr5e-05-3
63
+ [4]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr5e-05-4
64
+ [5]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr5e-05-5
65
+ [6]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr3e-05-1
66
+ [7]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr3e-05-2
67
+ [8]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr3e-05-3
68
+ [9]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr3e-05-4
69
+ [10]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr3e-05-5
70
+ [11]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr5e-05-1
71
+ [12]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr5e-05-2
72
+ [13]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr5e-05-3
73
+ [14]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr5e-05-4
74
+ [15]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr5e-05-5
75
+ [16]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr3e-05-1
76
+ [17]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr3e-05-2
77
+ [18]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr3e-05-3
78
+ [19]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr3e-05-4
79
+ [20]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr3e-05-5
80
+
81
+ The result in bold shows the performance of the current viewed model.
82
+
83
+ Additionally, the Flair [training log](training.log) and [TensorBoard logs](../../tensorboard) are also uploaded to the model
84
+ hub.