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.

Model Details

Base Model: Falconsai/text_summarization (T5-small)
Fine-tuned by: Wordcab Team
Model type: T5 (Text-to-Text Transfer Transformer)
Language: English
License: Apache 2.0
Demo: Try it on Spaces

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 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: Falconsai/text_summarization
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

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%
Successfully differentiates brief vs full summaries (2.5x length difference)
Robust 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 Falconsai/text_summarization T5 model
Fine-tuned on argilla/FinePersonas-Conversations-Email-Summaries dataset
Training performed using HuggingFace Transformers
Special thanks to the open-source community

Contact

For questions or feedback, please open an issue on the model repository.