Practical overview
When it fits
Use it when you have tokenized source-summary pairs and want a code-based workflow for training, decoding and ROUGE evaluation rather than a hosted summarization service.
Who it is for
- Developers adapting a research-style summarization training pipeline
- Researchers working with tokenized source-summary datasets
- Teams interested in pointer-generator, attention, coverage, or reinforcement-learning summarization experiments
Documented uses
- Train an abstractive summarization model on paired source and reference-summary text.
- Prepare Google sentence compression or CNN/Daily Mail data using included scripts.
- Decode summaries with beam search and assess them using ROUGE.
- Resume a stopped training run from a saved training state.
Requirements & limits
The repository declares an MIT license.
- The README specifies Python 3.6 and PyTorch 0.4.0, as well as NumPy, tqdm and matplotlib.
- Input data must already be tokenized, with source and summary separated by a tab; paragraph breaks use .
- NLTK is required for one preprocessing script, not identified as a universal requirement.
- ROUGE scoring requires ROUGE-1.5.5 and its data directory in the project’s data/ directory; pyrouge is not required.
- The model’s pointer-generator option requires encoder attention, according to the model parameter documentation.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Prepare data and run summarization training
Prepare line-oriented source-summary examples and start training with the repository’s configured parameters.
Before you start
- Python 3.6 and PyTorch 0.4.0, as specified by the README
- NumPy, tqdm and matplotlib
- A text or gzipped text file with one example per line; source and summary separated by a tab, both already tokenized
Steps
Format each example as a source text and summary separated by a tab on one line; use for paragraph breaks.
Place the prepared data in the format expected by the project’s data workflow.
Run train.py to start training with parameters from params.py; use command-line options to override parameters or --resume_from X.train.pt to resume a saved training state.
Expected outcome
Training starts using the configured data and parameters, or resumes from the specified saved state.
Guide limitations
- The README does not specify the required dataset path or provide a complete command line for selecting an input file.
- One preprocessing script requires NLTK for tokenization; the training data itself must already be tokenized.
Evaluate a trained model with ROUGE
Run the project’s test script to score a trained model and optionally save decoded summaries.
Before you start
- A trained model available as the latest model for the parameters in params.py
- ROUGE-1.5.5.pl and its data directory placed under data/ so that a data/data/ directory exists
Steps
Place ROUGE-1.5.5.pl and its data directory under the project’s data/ directory.
Run test.py to evaluate the latest model trained using the parameters in params.py.
Use the test script’s option to save decoded summaries if desired.
Expected outcome
The script performs beam-search decoding and prints ROUGE scores; decoded summaries can also be saved.
Guide limitations
- The README does not specify the command-line option for saving decoded summaries.
- The test script evaluates the latest model associated with params.py.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 18 October 2026
- Original website title
- ymfa/seq2seq-summarizer · Official homepage title · retained in its source language