--- language: de license: mit tags: - flair - token-classification - sequence-tagger-model - hetzner - hetzner-gex44 - hetzner-gpu base_model: google-bert/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 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: [`16`, `8`] * 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 GX44](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` | [0.9346][1] | [**0.9388**][2] | [0.9301][3] | [0.9291][4] | [0.9346][5] | 0.9334 ± 0.0039 | | `bs16-e10-lr5e-05` | [0.9316][6] | [0.9328][7] | [0.9341][8] | [0.9315][9] | [0.9248][10] | 0.931 ± 0.0036 | | `bs8-e10-lr3e-05` | [0.9234][11] | [0.9391][12] | [0.9207][13] | [0.9191][14] | [0.9394][15] | 0.9283 ± 0.0101 | | `bs16-e10-lr3e-05` | [0.9136][16] | [0.9269][17] | [0.9231][18] | [0.9251][19] | [0.9247][20] | 0.9227 ± 0.0053 | [1]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs8-e10-lr5e-05-1 [2]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs8-e10-lr5e-05-2 [3]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs8-e10-lr5e-05-3 [4]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs8-e10-lr5e-05-4 [5]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs8-e10-lr5e-05-5 [6]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs16-e10-lr5e-05-1 [7]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs16-e10-lr5e-05-2 [8]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs16-e10-lr5e-05-3 [9]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs16-e10-lr5e-05-4 [10]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs16-e10-lr5e-05-5 [11]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs8-e10-lr3e-05-1 [12]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs8-e10-lr3e-05-2 [13]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs8-e10-lr3e-05-3 [14]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs8-e10-lr3e-05-4 [15]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs8-e10-lr3e-05-5 [16]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs16-e10-lr3e-05-1 [17]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs16-e10-lr3e-05-2 [18]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs16-e10-lr3e-05-3 [19]: https://hf.co/stefan-it/flair-co-funer-german_bert_base-bs16-e10-lr3e-05-4 [20]: https://hf.co/stefan-it/flair-co-funer-german_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.