th-nuernberg/gbert-large-german-counseling-gecco

This model is a fine-tuned version of deepset/gbert-large trained with the German E-Counseling Conversation Dataset, created at the Technische Hochschule Nürnberg (see github.com/th-nuernberg/gecco-dataset).

It achieves the following results on the evaluation set: Accuracy 0.78, F1 0.66.

Contact:

• Prof. Dr. Jens Albrecht
• Prof. Dr. Robert Lehmann

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 64
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 16

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1
3.3924	1.0	20	2.9410	0.2032	0.0418
2.7028	2.0	40	2.2499	0.4806	0.2366
2.0665	3.0	60	1.7404	0.6129	0.3537
1.5	4.0	80	1.3602	0.6839	0.4109
1.0794	5.0	100	1.1377	0.7355	0.4971
0.7965	6.0	120	1.0123	0.7548	0.5518
0.6438	7.0	140	0.9806	0.7613	0.5547
0.5039	8.0	160	0.9452	0.7742	0.6019
0.4058	9.0	180	0.9218	0.7774	0.5907
0.3363	10.0	200	0.9373	0.7710	0.6157
0.2451	11.0	220	0.9751	0.7548	0.5955
0.1997	12.0	240	0.9197	0.7839	0.6526
0.1765	13.0	260	0.9187	0.7806	0.6425
0.1453	14.0	280	0.9431	0.7742	0.6357
0.1216	15.0	300	0.9388	0.7839	0.6534
0.1097	16.0	320	0.9290	0.7839	0.6645

Framework versions

• Transformers 4.35.1
• Pytorch 1.10.1+cu111
• Datasets 2.14.7
• Tokenizers 0.14.1