|
|
--- |
|
|
license: mit |
|
|
pipeline_tag: text-classification |
|
|
--- |
|
|
## Roberta for Justification analyst |
|
|
|
|
|
This model is a fine-tuned version of the Roberta architecture that has been trained specifically for sequence classification. The fine-tuning process involved using the PyTorch deep learning framework and specific hyperparameters (2-4e, 1-8 epsilon) with Adagrad optimizer. |
|
|
|
|
|
--- |
|
|
## Example Usage |
|
|
|
|
|
To use the model, first load it in PyTorch: |
|
|
|
|
|
```python |
|
|
import torch |
|
|
from transformers import RobertaForSequenceClassification, RobertaTokenizer |
|
|
# Load the fine-tuned model |
|
|
model = RobertaForSequenceClassification.from_pretrained('Dzeniks/justification-analyst') |
|
|
|
|
|
# Load the tokenizer |
|
|
tokenizer = RobertaTokenizer.from_pretrained('Dzeniks/justification-analyst') |
|
|
|
|
|
# Tokenize the input sequence |
|
|
input_text = "This is a sample input sequence" |
|
|
input = tokenizer.encode_plus(claim, evidence, return_tensors="pt") |
|
|
# Use the model to make a prediction |
|
|
model.eval() |
|
|
with torch.no_grad(): |
|
|
prediction = model(**x) |
|
|
predictions = torch.argmax(outputs[0], dim=1).item() |
|
|
``` |
|
|
|
|
|
## Classification Labels |
|
|
|
|
|
The model was trained on a dataset consisting of claims and evidence, where the goal was to classify each claim as either supporting, refuting, or not having enough information to make a decision. The labels used for this task are as follows: |
|
|
|
|
|
- Label 0: Supports |
|
|
- Label 1: Refutes |
|
|
- Label 2: Not enough information |
|
|
|
