moz-v/samsum-pg-checkpoints

SAMSum pointer-generator checkpoints

This repository contains the Task 2 checkpoints for a from-scratch pointer-generator model trained on knkarthick/samsum.

Files

• pg_nocov.pt: pointer-generator without coverage. The checkpoint is a dictionary with a state_dict key.
• pg_cov.pt: pointer-generator with coverage. The checkpoint is a dictionary with a state_dict key.
• vocab.json: token list used by both checkpoints.

Architecture and training

• BiLSTM encoder with 256 hidden units per direction.
• LSTM decoder with 256 hidden units.
• Bahdanau attention.
• 128-dimensional embeddings trained from scratch.
• Vocabulary size 30,586.
• Adagrad with learning rate 0.15 and initial accumulator 0.1.
• Gradient clipping at 2.0.
• Beam search width 4.
• Coverage loss weight 1.0 for the final 3,000 iterations.

Reported test metrics

Variant	ROUGE-1	ROUGE-2	ROUGE-L	Compression	Extractive density
No coverage	0.3622	0.1454	0.3056	0.2125	0.8392
Coverage	0.3661	0.1323	0.2997	0.1995	0.8785

ROUGE is computed with the same unstemmed implementation used in Task 1.

Usage

from huggingface_hub import hf_hub_download

repo_id = "moz-v/samsum-pg-checkpoints"
for filename in ["pg_nocov.pt", "pg_cov.pt", "vocab.json"]:
    hf_hub_download(repo_id=repo_id, filename=filename, local_dir="task2_33k")

License

The SAMSum dataset is distributed under CC-BY-NC-ND-4.0. These checkpoints are provided for coursework use only. The source code is not distributed in this repository.