hebert-finetuned-hebrew-metaphor
The model is fine-tuned to determine if a word in a sentence is used metaphorically or literally.
The model was trained for the following verbs: לחלום, לחתוך, לעוף, לפרק, להדליק, לכבס, לכופף, לרסק, לבשל, למחוק, לקפוץ, לקרוע, לקצור, לרקוד, לשבור, לשדוד, לשתות, לטחון, לתפור, לזרוע
This model is a fine-tuned version of avichr/heBERT on HebrewMetaphors dataset. It achieves the following results on the evaluation set:
- Loss: 0.4682
- Accuracy: 0.9510
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 15
Training results
Training Loss | Epoch | Step | Validation Loss | Accuracy |
---|---|---|---|---|
No log | 1.0 | 389 | 0.1813 | 0.9379 |
0.2546 | 2.0 | 778 | 0.2309 | 0.9479 |
0.08 | 3.0 | 1167 | 0.3342 | 0.9492 |
0.0298 | 4.0 | 1556 | 0.4076 | 0.9460 |
0.0298 | 5.0 | 1945 | 0.3803 | 0.9485 |
0.0105 | 6.0 | 2334 | 0.3674 | 0.9454 |
0.0077 | 7.0 | 2723 | 0.5356 | 0.9410 |
0.0088 | 8.0 | 3112 | 0.4776 | 0.9422 |
0.0044 | 9.0 | 3501 | 0.4258 | 0.9504 |
0.0044 | 10.0 | 3890 | 0.4305 | 0.9523 |
0.001 | 11.0 | 4279 | 0.4357 | 0.9548 |
0.0031 | 12.0 | 4668 | 0.4770 | 0.9473 |
0.0015 | 13.0 | 5057 | 0.4604 | 0.9523 |
0.0015 | 14.0 | 5446 | 0.4670 | 0.9510 |
0.0022 | 15.0 | 5835 | 0.4682 | 0.9510 |
Framework versions
- Transformers 4.30.2
- Pytorch 2.0.1+cu118
- Datasets 2.13.1
- Tokenizers 0.13.3
About Us
Created by Doron Ben-chorin, Matan Ben-chorin, Tomer Tzipori, Guided by Dr. Oren Mishali. This is our project as part of computer engineering studies in the Faculty of Electrical Engineering combined with Computer Science at Technion, Israel Institute of Technology. For more cooperation, please contact email:
Doron Ben-chorin: doronbh7@gmail.com
Matan Ben-chorin: matan.bh1@gmail.com
Tomer Tzipori: TomerTzipori@gmail.com
- Downloads last month
- 11,005