hulat-group/alignscore-base-checkpoint

AlignScore Checkpoints

This repository provides the AlignScore checkpoints required to run AlignScore-based evaluation within HULAT's NLP evaluation pipelines.

AlignScore is an automatic evaluation metric designed to assess the alignment between a source text and a generated text, supporting tasks such as factual consistency evaluation, semantic preservation, summarization evaluation, and text simplification evaluation.

This repository currently includes:

• AlignScore-base.ckpt
• AlignScore-large.ckpt

These files are intended to be used as pretrained checkpoints for AlignScore, especially in evaluation workflows involving generated or simplified texts.

---

Repository Contents

File	Description
AlignScore-base.ckpt	Base AlignScore checkpoint
AlignScore-large.ckpt	Large AlignScore checkpoint
.gitattributes	Git LFS configuration
README.md	Repository documentation

Important: These files are checkpoints for the AlignScore metric. They are not standalone text generation models.

---

Intended Use

These checkpoints are intended for research and evaluation purposes, particularly for:

• Semantic preservation evaluation
• Factual consistency evaluation
• Text simplification evaluation
• Summarization evaluation
• Evaluation of generated text against a source document
• Integration into NLP benchmark pipelines

Within HULAT, these checkpoints are used as part of broader evaluation workflows for text simplification and accessibility-oriented NLP research.

---

Important Notes

English AlignScore checkpoint

For the English version of AlignScore, the checkpoint must be downloaded manually from this repository and placed in the expected local path.

Manual download:

https://huggingface.co/hulat-group/alignscore-base-checkpoint/tree/main

How to Use

These checkpoints should be loaded through an AlignScore implementation or a compatible evaluation pipeline.

Example conceptual usage:

from alignscore import AlignScore

scorer = AlignScore(
    model="roberta-base",
    batch_size=16,
    device="cuda:0",
    ckpt_path="AlignScore_v2_es/checkpoint/checkpoints/AlignScore-base.ckpt"
)

score = scorer.score(
    contexts=["The original source text goes here."],
    claims=["The generated or simplified text goes here."]
)

print(score)