text-summarizer-t5

Text Summarizer — T5-small

A fine-tuned T5-small model for text summarization, trained on a subset of the SAMSum conversational dataset.

Model Details

• Base model: t5-small
• Task: Text summarization
• Architecture: T5 encoder-decoder Transformer
• Parameters: ~60.5M
• Training dataset: SAMSum
• Training samples: 4,000
• Validation samples: 500
• Epochs: 3
• Batch size: 3
• Weight decay: 0.01
• Warmup steps: 400
• Training hardware: NVIDIA Tesla T4
• Final training loss: ~0.762

Intended Use

This model is intended for experimentation and learning around Transformer-based text summarization.

It works primarily with conversational text similar to the SAMSum dataset used during fine-tuning.

Limitations

This model was fine-tuned on a relatively small subset of the SAMSum dataset, so summary quality can vary depending on the input.

It should not be considered a general-purpose summarization model, particularly for long or highly technical documents.

Training

The model was fine-tuned using:

• Hugging Face Transformers
• PyTorch
• T5-small
• Hugging Face Trainer

The training and experimentation code is available in the project repository.

Project Repository:
https://github.com/AlishaShaikh20/text-summarizer

Live Demo:
https://alishashaikh20--text-summariser-fastapi-app.modal.run/

Example

Input

Sarah: Hey, are we still meeting tomorrow for the project?
John: Yes, I think 10 AM works.
Sarah: Perfect. Should we meet at the library?
John: The library is fine. I'll bring the project notes.

Generated Summary

Sarah and John are meeting tomorrow at 10 AM for the project. John will bring the project notes.

Technical Stack

• Python
• PyTorch
• Hugging Face Transformers
• T5-small
• FastAPI
• Modal

Repository

The complete project, including the training notebook and FastAPI application, is available here:

https://github.com/AlishaShaikh20/text-summarizer

License

This model is shared for educational and portfolio purposes.

Please refer to the original t5-small and SAMSum dataset licenses for their respective usage terms.