File size: 4,291 Bytes
791d147
 
 
 
 
 
 
 
 
 
 
9a776b9
 
791d147
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9a776b9
 
791d147
 
40e2e3a
791d147
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
---
language: de
license: mit
tags:
- flair
- token-classification
- sequence-tagger-model
- hetzner
- hetzner-gex44
- hetzner-gpu
base_model: dbmdz/bert-base-german-cased
datasets:
- stefan-it/co-funer
widget:
- text: Wesentliche Tätigkeiten der Compliance-Funktion wurden an die Mercurtainment
    AG , Düsseldorf , ausgelagert .
---

# Fine-tuned Flair Model on CO-Fun NER Dataset

This Flair model was fine-tuned on the
[CO-Fun](https://arxiv.org/abs/2403.15322) NER Dataset using German DBMDZ BERT as backbone LM.

## Dataset

The [Company Outsourcing in Fund Prospectuses (CO-Fun) dataset](https://arxiv.org/abs/2403.15322) consists of
948 sentences with 5,969 named entity annotations, including 2,340 Outsourced Services, 2,024 Companies, 1,594 Locations
and 11 Software annotations.

Overall, the following named entities are annotated:

* `Auslagerung` (engl. outsourcing)
* `Unternehmen` (engl. company)
* `Ort` (engl. location)
* `Software`

## Fine-Tuning

The latest [Flair version](https://github.com/flairNLP/flair/tree/42ea3f6854eba04387c38045f160c18bdaac07dc) is used for
fine-tuning.

A hyper-parameter search over the following parameters with 5 different seeds per configuration is performed:

* Batch Sizes: [`8`, `16`]
* Learning Rates: [`5e-05`, `3e-05`]

More details can be found in this [repository](https://github.com/stefan-it/co-funer). All models are fine-tuned on a
[Hetzner GEX44](https://www.hetzner.com/dedicated-rootserver/matrix-gpu/) with an NVIDIA RTX 4000.

## Results

A hyper-parameter search with 5 different seeds per configuration is performed and micro F1-score on development set
is reported:

| Configuration      | Seed 1       | Seed 2       | Seed 3       | Seed 4           | Seed 5       | Average         |
|--------------------|--------------|--------------|--------------|------------------|--------------|-----------------|
| `bs8-e10-lr5e-05`  | [0.9378][1]  | [0.928][2]   | [0.9383][3]  | [0.9374][4]      | [0.9364][5]  | 0.9356 ± 0.0043 |
| `bs8-e10-lr3e-05`  | [0.9336][6]  | [0.9366][7]  | [0.9299][8]  | [0.9417][9]      | [0.9281][10] | 0.934 ± 0.0054  |
| `bs16-e10-lr5e-05` | [0.927][11]  | [0.9341][12] | [0.9372][13] | [0.9283][14]     | [0.9329][15] | 0.9319 ± 0.0042 |
| `bs16-e10-lr3e-05` | [0.9141][16] | [0.9321][17] | [0.9175][18] | [**0.9391**][19] | [0.9177][20] | 0.9241 ± 0.0109 |

[1]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr5e-05-1
[2]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr5e-05-2
[3]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr5e-05-3
[4]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr5e-05-4
[5]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr5e-05-5
[6]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr3e-05-1
[7]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr3e-05-2
[8]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr3e-05-3
[9]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr3e-05-4
[10]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs8-e10-lr3e-05-5
[11]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr5e-05-1
[12]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr5e-05-2
[13]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr5e-05-3
[14]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr5e-05-4
[15]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr5e-05-5
[16]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr3e-05-1
[17]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr3e-05-2
[18]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr3e-05-3
[19]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr3e-05-4
[20]: https://hf.co/stefan-it/flair-co-funer-german_dbmdz_bert_base-bs16-e10-lr3e-05-5

The result in bold shows the performance of the current viewed model.

Additionally, the Flair [training log](training.log) and [TensorBoard logs](../../tensorboard) are also uploaded to the model
hub.