bert-base-german-cased-defakts-fake-binary

This Model is finetuned for sequence classification (binary fake-news classification task) on the german DeFaktS-Dataset. It achieves the following results on the evaluation set:

• Loss: 0.3608
• Accuracy: 0.8531
• F1: 0.8392
• Precision: 0.8636
• Recall: 0.8283

Model description

This Model is finetuned for sequence classification

Dataset

Trained on the DeFactS dataset https://github.com/caisa-lab/DeFaktS-Dataset-Disinformaton-Detection, feature catposfake/catneutral to detect fake news

Intended uses & limitations

Fake news classification

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 6

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy@de	F1@de	Precision@de	Recall@de	Loss@de
0.5531	0.0888	50	0.4191	0.8071	0.7958	0.7993	0.7931	0.4192
0.4341	0.1776	100	0.3987	0.8186	0.8132	0.8100	0.8198	0.3988
0.3854	0.2664	150	0.3816	0.8206	0.8100	0.8138	0.8071	0.3817
0.4028	0.3552	200	0.3703	0.8301	0.8221	0.8218	0.8223	0.3704
0.3784	0.4440	250	0.3750	0.8276	0.8065	0.8491	0.7933	0.3752
0.3622	0.5329	300	0.3465	0.8461	0.8389	0.8384	0.8393	0.3465
0.3945	0.6217	350	0.4596	0.7706	0.7704	0.7934	0.8006	0.4595
0.4073	0.7105	400	0.3360	0.8531	0.8419	0.8549	0.8343	0.3361
0.3779	0.7993	450	0.3440	0.8451	0.8399	0.8366	0.8455	0.3441
0.3596	0.8881	500	0.3608	0.8531	0.8392	0.8636	0.8283	0.3610
0.3588	0.9769	550	0.3468	0.8516	0.8375	0.8620	0.8266	0.3468
0.287	1.0657	600	0.3416	0.8591	0.8527	0.8517	0.8539	0.3416
0.2395	1.1545	650	0.3976	0.8531	0.8419	0.8547	0.8345	0.3977
0.2278	1.2433	700	0.3635	0.8441	0.8387	0.8355	0.8438	0.3635
0.2495	1.3321	750	0.3294	0.8581	0.8518	0.8506	0.8530	0.3294
0.2455	1.4210	800	0.3448	0.8581	0.8516	0.8507	0.8526	0.3448
0.2472	1.5098	850	0.3743	0.8626	0.8527	0.8635	0.8460	0.3745

Framework versions

• Transformers 4.45.2
• Pytorch 2.3.1+cu121
• Tokenizers 0.20.3