---
license: mit
tags:
- text-generation
- customer-support
- gpt2
- fine-tuned
datasets:
- customer-support-on-twitter
metrics:
- bleu
- rouge
- bertscore
- perplexity
model-index:
- name: customer-support-gpt2
  results:
  - task:
      type: text-generation
      name: Customer Support Response Generation
    dataset:
      type: customer-support-on-twitter
      name: Customer Support on Twitter
    metrics:
    - type: bleu
      value: 0.0568
    - type: rouge-l
      value: 0.1344
    - type: bertscore-f1
      value: 0.7507
    - type: perplexity
      value: 17.86
---

# Customer Support GPT-2

Fine-tuned GPT-2 model for generating customer support responses on Twitter.

## Model Description

This model is a fine-tuned version of GPT-2 (124M parameters) trained on the Customer Support on Twitter dataset. It generates contextually appropriate customer support responses based on customer queries.

## Training Results

- **Perplexity**: 17.86 (down from ~35 baseline)
- **BLEU Score**: 0.0568
- **ROUGE-L**: 0.1344
- **BERTScore F1**: 0.7507

## Baseline Comparison

- **BLEU Improvement**: 5.1% over pre-trained GPT-2
- **ROUGE Improvement**: 165.2% over pre-trained GPT-2

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("suhasramanand/customer-support-gpt2")
tokenizer = AutoTokenizer.from_pretrained("suhasramanand/customer-support-gpt2")

def generate_response(customer_query):
    input_text = f"Customer: {customer_query}\nAgent:"
    inputs = tokenizer(input_text, return_tensors="pt")
    
    outputs = model.generate(
        **inputs,
        max_length=150,
        temperature=0.7,
        top_p=0.9,
        do_sample=True
    )
    
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response.split("Agent:")[-1].strip()

# Example
response = generate_response("I can't log into my account")
print(response)
```

## Training Details

- **Base Model**: GPT-2 (124M parameters)
- **Configuration**: Balanced
- **Training Samples**: 40,000 conversation pairs
- **Dataset**: Customer Support on Twitter (Kaggle)

## Dataset

Trained on the Customer Support on Twitter dataset containing over 2.8 million tweets and replies from various companies' customer support accounts.

## Limitations

- May generate generic responses for complex queries
- Performance varies by domain (works best for common support scenarios)
- Requires post-processing for production use

## Citation

```bibtex
@misc{customer-support-gpt2,
  title={Customer Support Response Generation with Fine-tuned GPT-2},
  author={Your Name},
  year={2024},
  url={https://huggingface.co/suhasramanand/customer-support-gpt2}
}
```