This model is a fine-tuned version of Leonard Konle's fiction-gbert.
It was fine-tuned for ten epochs on the Deutscher Roman Korpus (DROC) on literary character detection using a standard token-classification head. However, in deviation from most other models, this model detects named entities and nouns (matching both "Harry" and "Zauberer") referencing a character.
The model achieves a 92.12 / 89.98 % F1 score on the semi-official DROC validation and test sets.
The code to reproduce the dataset and training can be accessed via Github
Additional hyperparameters are:
- Num Epochs: 10
- Batch-Size: 8
- Optimizer: AdamW
- Learning rate: 2e-05
- Weight-Decay: 0.1
- Scheduler: Linear Warmup for the first 10 % of the training, with a linear decay for the remainder.
- Precision: 32bit
- Training-Framework: Trident
ID2Label-Map:
{
0: "O",
1: "B-PER",
2: "I-Per"
}