BERT Fine-Tuned on Winograd NLI
A fine-tuned BERT model using the Winograd NLI dataset.
Model Details
Description
This model is based on the BERT base (uncased) architecture and has been fine-tuned on the Winograd NLI dataset.
- Developed by: Cesar Gonzalez-Gutierrez
- Funded by: ERC
- Architecture: BERT-base
- Base model: BERT base model (uncased)
- Language: English
- License: Apache 2.0
Seed Initializations
Alternative models trained using different initialization seeds are available and can be accessed using specific branches:
| Random Seed | Branch |
|---|---|
| 120 | seed-120 |
| 220 | seed-220 |
| 320 | seed-320 |
| 420 | seed-420 |
| 520 | seed-520 |
To load a model from a specific branch, use the revision parameter:
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained("<model>", revision="seed-120")
Sources
[Information pending]
Training Details
Fine-tuning was performed end-to-end using a grid search over key hyperparameters. Model performance was evaluated based on validation loss computed on the development set. After identifying the optimal hyperparameter configuration, the final model was retrained on the entire training dataset.
Training Data
The model was trained on the Winograd NLI training split. Validation was performed using a random 20% subset of this training data. The original validation split was used as the test set, since the original test split does not include labels.
Training Hyperparameters
- Epochs: 1-4
- Batch size: {16, 32}
- Learning rate: {5e-5, 3e-5, 2e-5}
- Validation metric: loss
- Precision: fp16
Uses
This model can be used for classification tasks aligned with the structure and intent of the Winograd NLI dataset.
For broader guidance, refer to the BERT base model’s Inteded Uses & Limitations.
Bias, Risks, and Limitations
This model inherits the potential risks and limitations of its base model. For more details, refer to the Limitations and bias section of the original model documentation.
Additionally, it may reflect or amplify patterns and biases present in the Winograd NLI training data.
Hardware
- Hardware Type: NVIDIA Tesla V100 PCIE 32GB
- Cluster Provider: Artemisa
- Compute Region: EU
Citation
If you use this model in your research, please cite both the base BERT model and the Winograd NLI source.