Upload README.md

README.md

@@ -0,0 +1,124 @@
+
+# roberta-large-fallacy-classification
+
+This model is a fine-tuned version of `roberta-large` trained for logical fallacy detection on the [Logical Fallacy Dataset](https://huggingface.co/datasets/tasksource/logical-fallacy). It is capable of classifying various types of logical fallacies in text.
+
+## Model Details
+
+- **Base Model**: `roberta-large`
+- **Dataset**: Logical Fallacy Dataset
+- **Number of Classes**: 13
+- **Training Parameters**:
+  - **Learning Rate**: 5e-6 with cosine decay scheduler
+  - **Batch Size**: 8 (with gradient accumulation for effective batch size of 16)
+  - **Weight Decay**: 0.3
+  - **Label Smoothing**: 0.1
+  - **Mixed Precision (FP16)**: Enabled
+  - **Early Stopping**: Used with patience of 2 epochs
+- **Training Time**: Approximately 10 epochs
+
+## Example Pipeline
+
+To use the model for quick classification with a text pipeline:
+
+```python
+from transformers import pipeline
+
+# Replace with your Hugging Face model path
+model_path = "MidhunKanadan/roberta-large-fallacy-classification"
+
+# Initialize the text classification pipeline
+pipe = pipeline("text-classification", model=model_path, tokenizer=model_path, device=0)  # Set device=0 to use GPU if available
+
+# Define a sample text
+text = "The rooster crows always before the sun rises, therefore the crowing rooster causes the sun to rise."
+
+# Make a prediction
+result = pipe(text)
+
+# Print the predicted label and score
+print(f"Predicted Label: {result[0]['label']}")
+print(f"Score: {result[0]['score']:.4f}")
+```
+
+Expected Output:
+```
+Predicted Label: false causality
+Score: 0.8938
+```
+
+## Full Classification Example
+
+For more control, load the model and tokenizer directly and perform classification:
+
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+
+# Load your model and tokenizer from Hugging Face
+model = AutoModelForSequenceClassification.from_pretrained(model_path)
+tokenizer = AutoTokenizer.from_pretrained(model_path)
+model.to("cuda")  # Move to GPU if available
+
+# Define a sample text
+text = "The rooster crows always before the sun rises, therefore the crowing rooster causes the sun to rise."
+
+# Tokenize the input text
+inputs = tokenizer(text, return_tensors="pt").to("cuda")
+
+# Run the model and get logits
+with torch.no_grad():
+    logits = model(**inputs).logits
+
+# Apply softmax to get probabilities
+probabilities = torch.nn.functional.softmax(logits, dim=1)[0]
+
+# Print each label and its corresponding score
+for label, score in zip(model.config.id2label.values(), probabilities):
+    print(f"{label}: {score.item():.4f}")
+
+```
+
+Expected Output:
+```
+ad hominem: 0.0025
+appeal to emotion: 0.0037
+false dilemma: 0.0053
+false causality: 0.8938
+fallacy of relevance: 0.0059
+ad populum: 0.0053
+faulty generalization: 0.0104
+fallacy of credibility: 0.0040
+fallacy of extension: 0.0042
+intentional: 0.0036
+circular reasoning: 0.0127
+fallacy of logic: 0.0366
+equivocation: 0.0121
+```
+
+## Training Details
+
+The model was trained using the following parameters:
+- **Optimizer**: AdamW
+- **Learning Rate**: 5e-6 with cosine decay scheduler
+- **Batch Size**: 8 (with gradient accumulation to achieve effective batch size of 16)
+- **Weight Decay**: 0.3
+- **Label Smoothing Factor**: 0.1
+- **Early Stopping**: Enabled (patience = 2)
+- **Mixed Precision**: Enabled (FP16)
+
+## Dataset
+
+- **Dataset Name**: Logical Fallacy Dataset
+- **Source**: [Hugging Face Datasets](https://huggingface.co/datasets/tasksource/logical-fallacy)
+- **Number of Classes**: 14 fallacies (e.g., ad hominem, appeal to emotion, faulty generalization, etc.)
+
+## Limitations
+
+This model may not generalize well to all types of logical fallacies due to the limited size of the dataset and potential class imbalance. It may require additional fine-tuning or data augmentation to perform effectively in production.
+
+## Evaluation
+
+The model achieved the following evaluation metrics:
+- **Accuracy**: Varies by dataset split; see training logs for more details.
+- **F1 Score**: Varies by dataset split; see training logs for more details.