deberta-v3-large-sentiment
This model is a fine-tuned version of microsoft/deberta-v3-large on an tweet_eval dataset.
Model description
Test set results:
Model | Emotion | Hate | Irony | Offensive | Sentiment |
---|---|---|---|---|---|
deberta-v3-large | 86.3 | 61.3 | 87.1 | 86.4 | 73.9 |
BERTweet | 79.3 | - | 82.1 | 79.5 | 73.4 |
RoB-RT | 79.5 | 52.3 | 61.7 | 80.5 | 69.3 |
Intended uses & limitations
Classifying attributes of interest on tweeter like data.
Training and evaluation data
tweet_eval dataset.
Training procedure
Fine tuned and evaluated with run_glue.py
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 7e-06
- 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
- lr_scheduler_warmup_steps: 50
- num_epochs: 10.0
- label_smoothing_factor: 0.1
Training results
Training Loss | Epoch | Step | Validation Loss | Accuracy |
---|---|---|---|---|
1.2787 | 0.49 | 100 | 1.1127 | 0.4866 |
1.089 | 0.98 | 200 | 0.9668 | 0.7139 |
0.9134 | 1.47 | 300 | 0.8720 | 0.7834 |
0.8618 | 1.96 | 400 | 0.7726 | 0.7941 |
0.686 | 2.45 | 500 | 0.7337 | 0.8209 |
0.6333 | 2.94 | 600 | 0.7350 | 0.8235 |
0.5765 | 3.43 | 700 | 0.7561 | 0.8235 |
0.5502 | 3.92 | 800 | 0.7273 | 0.8476 |
0.5049 | 4.41 | 900 | 0.8137 | 0.8102 |
0.4695 | 4.9 | 1000 | 0.7581 | 0.8289 |
0.4657 | 5.39 | 1100 | 0.8404 | 0.8048 |
0.4549 | 5.88 | 1200 | 0.7800 | 0.8369 |
0.4305 | 6.37 | 1300 | 0.8575 | 0.8235 |
0.4209 | 6.86 | 1400 | 0.8572 | 0.8102 |
0.3983 | 7.35 | 1500 | 0.8392 | 0.8316 |
0.4139 | 7.84 | 1600 | 0.8152 | 0.8209 |
0.393 | 8.33 | 1700 | 0.8261 | 0.8289 |
0.3979 | 8.82 | 1800 | 0.8328 | 0.8235 |
0.3928 | 9.31 | 1900 | 0.8364 | 0.8209 |
0.3848 | 9.8 | 2000 | 0.8322 | 0.8235 |
Framework versions
- Transformers 4.20.0.dev0
- Pytorch 1.9.0
- Datasets 2.2.2
- Tokenizers 0.11.6
- Downloads last month
- 22
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.