Abstractive Text Summarization Transformer

Responsible-use notice

Generated summaries may omit context, distort facts, or hallucinate. Do not enter confidential, sensitive, copyrighted, or personally identifiable text. Human review is required, especially for high-stakes or official use.

Interactive demo

Summarization workspace

Model not loaded

1. Source text

Paste a news article, report, complaint narrative, or other English text.

Example

0 words 0 characters — model tokens

Tokenized input preview

Load the model and generate a summary to view the first token IDs.

2. Generation controls

Experiment with decoding and output-length settings.

Runtime The browser uses full-precision ONNX files to avoid the incompatible quantized DistilBART decoder graphs. Minimum new tokens 30 Maximum new tokens 120 Beam count 4 More beams can improve search quality but increase latency. Length penalty 2.0 No-repeat n-gram 3 Long-document chunking Summarize token-aware chunks, then combine them in a second pass.

Model output

Generated summary

Your generated summary will appear here. The model runs only after you select Generate summary.

Inference latency —

Compression ratio —

Summary words —

Input chunks —

Runtime —

Model precision —

Architecture

Where the Transformer runs

Source text

Browser preprocessing

T5 encoder

Transformer decoder

Summary + metrics

The live application uses a browser-compatible fine-tuned T5-small ONNX checkpoint. The repository's separate Python training and benchmark layer uses DistilBART. Both are encoder-decoder Transformers, and no article is sent to a paid inference endpoint.

Python engineering layer

Full ML project preserved on GitHub

PyTorch and Hugging Face Transformers inference
ROUGE, BERTScore, latency, and compression evaluation
Lead-3, TextRank, and LSTM Seq2Seq comparison framework
Training, tests, notebooks, model card, and error analysis

Model evaluation

Publish only measured results

Loading evaluation status…

The committed evaluation file contains null values until the Python evaluation script is executed. This prevents fabricated portfolio metrics.

Model comparison

Modern Transformers vs LSTM Seq2Seq

Dimension	LSTM Seq2Seq with attention	Transformer project
Sequence processing	Recurrent, token by token	Attention-based parallel encoding
Long-range context	Compressed through recurrent states	Direct token-to-token attention
Pretraining	Earlier custom project	Pretrained T5 browser model and DistilBART Python model
Browser deployment	Requires a separate conversion	Fine-tuned T5-small ONNX runs through Transformers.js
Published metrics	Requires actual prior predictions	Generated for the Python DistilBART benchmark

Limitations

What reviewers should know

The initial full-precision model download is several hundred megabytes and can take time on a slower connection.
Browser speed depends on CPU performance, available memory, connection speed, and beam count.
The model is news-oriented and may perform less reliably on legal, medical, or specialized technical text.
Chunked long-document summarization can lose cross-section context or over-compress details.
Generated text may contain unsupported or incorrect facts and always requires review.

Abstractive Text Summarization