README.md

# T5 Email Summarizer - Brief & Full

## Model Description

This is a fine-tuned T5-small model specialized for email summarization. The model can generate both brief (one-line) and detailed (comprehensive) summaries of emails, and is robust to messy, informal inputs with typos and abbreviations.

### Key Features
- **Dual-mode summarization**: Supports both `summarize_brief:` and `summarize_full:` prefixes
- **Robust to informal text**: Handles typos, abbreviations, and casual language
- **Fast inference**: Based on T5-small (60M parameters)
- **Production-ready**: Optimized for real-world email processing

## Intended Uses & Limitations

### Intended Uses
- Summarizing professional and casual emails
- Email inbox management and prioritization
- Meeting invitation processing
- Customer support email triage
- Newsletter summarization

### Limitations
- Maximum input length: 512 tokens (~2500 characters)
- English language only
- May miss nuanced context in very long email threads
- Not suitable for legal or medical document summarization

## Training Data

The model was fine-tuned on the [argilla/FinePersonas-Conversations-Email-Summaries](https://huggingface.co/datasets/argilla/FinePersonas-Conversations-Email-Summaries) dataset containing 364,000 email-summary pairs with:
- Professional business emails
- Casual informal communications
- Meeting invitations and updates
- Technical discussions
- Customer service interactions

Training data was augmented with:
- Subtle typos and misspellings (50% of examples)
- Common abbreviations (pls, thx, tmrw, etc.)
- Various formatting styles

## Training Procedure

### Training Details
- **Base model**: google/t5-v1_1-small
- **Training epochs**: 1
- **Batch size**: 64
- **Learning rate**: 3e-4
- **Validation split**: 1%
- **Hardware**: NVIDIA GPU with mixed precision (fp16)
- **Framework**: HuggingFace Transformers 4.36.0

### Preprocessing
- Emails formatted as: `Subject: [subject]. Body: [content]`
- Two training examples created per email (brief and full summaries)
- Data augmentation applied to 50% of examples

## How to Use

### Installation
```bash
pip install transformers torch
```

### Basic Usage
```python
from transformers import T5ForConditionalGeneration, T5Tokenizer

# Load model and tokenizer
tokenizer = T5Tokenizer.from_pretrained("Wordcab/t5-small-email-summarizer")
model = T5ForConditionalGeneration.from_pretrained("Wordcab/t5-small-email-summarizer")

# Example email
email = """Subject: Team Meeting Tomorrow. Body: Hi everyone, 
Just a reminder that we have our weekly team meeting tomorrow at 2 PM EST. 
Please prepare your status updates and any blockers you're facing. 
We'll also discuss the Q4 roadmap. Thanks!"""

# Generate brief summary
inputs = tokenizer(f"summarize_brief: {email}", return_tensors="pt", max_length=512, truncation=True)
outputs = model.generate(**inputs, max_length=50, num_beams=2)
brief_summary = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(f"Brief: {brief_summary}")

# Generate full summary
inputs = tokenizer(f"summarize_full: {email}", return_tensors="pt", max_length=512, truncation=True)
outputs = model.generate(**inputs, max_length=150, num_beams=2)
full_summary = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(f"Full: {full_summary}")
```

### Advanced Usage with Long Emails

For emails longer than 512 tokens, consider using chunking:

```python
def summarize_long_email(email, model, tokenizer, mode="brief"):
    # Check if email fits in context
    tokens = tokenizer.encode(email)
    if len(tokens) <= 500:  # Leave room for prefix
        # Direct summarization
        prefix = f"summarize_{mode}:" if mode in ["brief", "full"] else "summarize:"
        inputs = tokenizer(f"{prefix} {email}", return_tensors="pt", max_length=512, truncation=True)
        outputs = model.generate(**inputs, max_length=150 if mode == "full" else 50)
        return tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # For longer emails, use strategic truncation or chunking
    # ... implement chunking strategy
```

## Performance Metrics

### Evaluation Results
- **ROUGE-L Score**: 0.42
- **Average inference time**: 0.63s (brief), 0.71s (full) on T4 GPU
- **Coherence score on messy inputs**: 80%

### Comparison with Base Model
- 8.3% improvement in quality over base Falconsai/text_summarization
- Successfully differentiates brief vs full summaries (2.5x length difference)
- Better handling of informal text and typos

## Deployment

### Using HuggingFace Inference API
```python
import requests

API_URL = "https://api-inference.huggingface.co/models/Wordcab/t5-small-email-summarizer"
headers = {"Authorization": "Bearer YOUR_HF_TOKEN"}

def query(payload):
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()

output = query({
    "inputs": "summarize_brief: Subject: Meeting. Body: Let's meet tomorrow at 3pm to discuss the project.",
})
```

### Using Text Generation Inference (TGI)
```bash
docker run --gpus all -p 8080:80 \
  -v t5-small-email-summarizer:/model \
  ghcr.io/huggingface/text-generation-inference:latest \
  --model-id Wordcab/t5-small-email-summarizer \
  --max-input-length 512 \
  --max-total-tokens 662
```

## Model Architecture

```
T5ForConditionalGeneration(
  (shared): Embedding(32128, 512)
  (encoder): T5Stack(
    (embed_tokens): Embedding(32128, 512)
    (block): ModuleList(
      (0-5): T5Block(...)
    )
    (final_layer_norm): T5LayerNorm()
    (dropout): Dropout(p=0.1)
  )
  (decoder): T5Stack(...)
  (lm_head): Linear(in_features=512, out_features=32128, bias=False)
)
```

## Citation

If you use this model, please cite:

```bibtex
@misc{wordcab2025t5email,
  title={T5 Email Summarizer - Brief & Full},
  author={Wordcab Team},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/Wordcab/t5-small-email-summarizer}
}
```

## License

This model is released under the Apache 2.0 License.

## Acknowledgments

- Based on Google's T5 architecture
- Fine-tuned using HuggingFace Transformers
- Training data derived from public email datasets
- Special thanks to the open-source community

## Contact

For questions or feedback, please open an issue on the [model repository](https://huggingface.co/Wordcab/t5-small-email-summarizer/discussions).