gbert-large-coherence_evaluation

Text classification model for coherence evaluation in German scientific texts

gbert-large-coherence_evaluation is a sequence classification model in the scientific domain in German, finetuned from the model gbert-large. It was trained using a custom annotated dataset of around 12,000 training and 3,000 test examples containing coherent and incoherent text sequences from wikipedia articles in german.

Compared to the base version, this model achieved a slightly higher peak accuracy (95.30%) on the validation set, observed at epoch 7. However, the base model reached its lowest evaluation loss (0.2347) earlier during training, suggesting that it converges faster but may underperform slightly in terms of generalization. These findings can inform future model selection depending on whether inference efficiency or accuracy is prioritized.

Training

Training was conducted on a 10 epoch fine-tuning approach:

Training was conducted using a standard Text classification objective. The model achieves an accuracy of approximately 95% on the evaluation set.

Here are the overall final metrics on the test dataset after 7 epochs of training:

• Accuracy: 0.9530156614461847
• Loss: 0.3258904218673706