English Split and Rephrase Gate

This is a binary English sentence-processing model. It predicts whether the current sentence should be split again before the rewriter is called.

Labels

The default pipeline threshold is 0.21 on the probability of continue. The threshold is stored in threshold.json for reproducible use with the companion code repository.

Intended use

Use this model as the decision step in an English split-and-rephrase pipeline. It can also be used independently when an application needs a stop/continue prediction for sentence decomposition.

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_id = "soundstarrain/english-split-and-rephrase-gate"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

text = "The service stores the file and returns a warning when validation fails."
inputs = tokenizer(text, return_tensors="pt", truncation=True)
with torch.inference_mode():
    probabilities = torch.softmax(model(**inputs).logits, dim=-1)[0]
label = model.config.id2label[int(probabilities.argmax())]
print(label, float(probabilities[1]))

The recommended end-to-end entry point is the companion GitHub repository, which loads this model together with the rewriter model and handles recursive processing and output checks.

Project repositories

Evaluation

On the held-out BiSECT test split, this checkpoint reached macro-F1 0.9979 with the validation-selected threshold 0.21. The metric describes the BiSECT split-and-rephrase task; performance on other English domains is not measured.

Limitations

The reported evaluation covers the BiSECT test split. Performance on other genres, languages, and sentence-processing tasks has not been measured.

Training data

The checkpoint was fine-tuned on filtered English split-and-rephrase examples derived from BiSECT. The training corpus is not included in this repository. The model card records GEM/BiSECT and the original paper for attribution.

License and attribution

Apache-2.0 applies to the released model files and metadata. The checkpoint is fine-tuned from answerdotai/ModernBERT-base. The training corpus is not redistributed here. See NOTICE.md for attribution.

Citation

Kim, J., Maddela, M., Kriz, R., Xu, W., and Callison-Burch, C. “BiSECT: Learning to Split and Rephrase Sentences with Bitexts.” EMNLP 2021.