# Model Details

**Model Name:** Fine-Tuned BART for Customer Support Resolution Generation

**Base Model:** facebook/bart-base

**Dataset:** bitext/Bitext-customer-support-llm-chatbot-training-dataset


**Training Device:** CUDA (GPU)

---

# Dataset Information

**Dataset Structure:**
DatasetDict({
train: Dataset({
features: ['input_text', 'target_text'],
num_rows: 24184
})
validation: Dataset({
features: ['input_text', 'target_text'],
num_rows: 2688
})
})


**Available Splits:**

- **Train:** 24,184 examples
- **Validation:** 2,688 examples

**Feature Representation:**

- **input_text:** Customer issue text (e.g., "Customer: How do I cancel my order?")
- **target_text:** Resolution text (e.g., "Log into the portal and cancel it there.")

---

# Training Details

**Training Process:**

- Fine-tuned for 3 epochs
- Loss reduced progressively across epochs

**Hyperparameters:**

- Epochs: 3
- Learning Rate: 2e-5
- Batch Size: 8
- Weight Decay: 0.01
- Mixed Precision: FP16

**Performance Metrics:**

- Final Training Loss: ~0.0140
- Final Validation Loss: ~0.0121


---

# Inference Example

```python
import torch
from transformers import BartForConditionalGeneration, BartTokenizer

def load_model(model_path):
    tokenizer = BartTokenizer.from_pretrained(model_path)
    model = BartForConditionalGeneration.from_pretrained(model_path).half()  # FP16
    model.eval()
    return model, tokenizer

def generate_resolution(issue, model, tokenizer, device="cuda"):
    input_text = f"Customer: {issue}"
    inputs = tokenizer(
        input_text,
        max_length=512,
        padding="max_length",
        truncation=True,
        return_tensors="pt"
    ).to(device)
    outputs = model.generate(
        inputs["input_ids"],
        max_length=128,
        num_beams=4,
        early_stopping=True
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Example usage
if __name__ == "__main__":
    model_path = "your-username/bart-resolution-summarizer-fp16"  # Replace with your HF repo
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model, tokenizer = load_model(model_path)
    model.to(device)

    issue = "How do I cancel my order?"
    resolution = generate_resolution(issue, model, tokenizer, device)
    print(f"Issue: {issue}")
    print(f"Resolution: {resolution}")
Expected Output:


Issue: How do I cancel my order?
Resolution: Log into the portal and cancel it there.
```

# Limitations
Model may struggle with issues requiring specific resolutions not well-represented in the training data (e.g., time-related queries like "When can I call support?").
Resolution extraction relied on heuristics, potentially missing nuanced answers in verbose responses.
# Future Improvements
Refine resolution extraction with more advanced NLP techniques or manual curation.
Fine-tune on additional customer support datasets for broader coverage.