Dataset Card for ccnews-embeddings-dim1024

This dataset contains precomputed 1024-dimensional embeddings of English news articles from the CCNEWS dataset using the sentence-transformers/all-roberta-large-v1 model. It also includes an out-of-distribution (OOD) subset of embeddings from Yahoo Answers, processed in the same way.

Note: This repository only provides the embeddings. If you are interested in the original text-embedding pairs for the CCNEWS subset, they are available in the related repository: ScarlettMagdaleno/ccnews_all-roberta-large-v1_1024.

Dataset Details

Dataset Description

The dataset is intended for tasks related to similarity search and evaluation under domain shift (in-distribution vs. out-of-distribution). It contains two main subsets:

• data/: Contains 614,664 embeddings (shape: [614664, 1024]) derived from the CCNEWS dataset.
• ood/: Contains 11,000 embeddings (shape: [11000, 1024]) derived from the Yahoo Answers dataset.

All embeddings were computed using the sentence-transformers/all-roberta-large-v1 model and stored in PyTorch .pt format.

• Curated by: Scarlett Magdaleno
• Language(s) (NLP): English
• License: Other (original datasets were redistributed under their respective licenses)

Dataset Sources

• CCNEWS: https://huggingface.co/datasets/sentence-transformers/ccnews
• Yahoo Answers: https://huggingface.co/datasets/sentence-transformers/yahoo-answers

Dataset Creation

Curation Rationale

The embeddings were generated to facilitate efficient similarity search and enable controlled experiments on generalization across domains, using CCNEWS as the in-distribution (ID) source and Yahoo Answers as the out-of-distribution (OOD) set.

Source Data

Data Collection and Processing

• Embeddings were computed using the sentence-transformers/all-roberta-large-v1 model.
• Only the text field from each dataset was used as input to the encoder.
• The outputs were saved in PyTorch format using torch.save.

Who are the source data producers?

• The original texts were produced by internet users and journalists.

• Embedding generation and formatting were done by Scarlett Magdaleno.

• Repository: https://huggingface.co/datasets/ScarlettMagdaleno/ccnews-embeddings-dim1024

Uses

Direct Use

• Evaluation of similarity search algorithms
• Benchmarking nearest neighbor methods under distributional shift

Out-of-Scope Use

• This dataset does not contain human-readable text or labels.

Dataset Structure

• Format: PyTorch .pt files.
• Each file contains a single torch.Tensor of shape [N, 1024], where N is the number of sentences encoded.

Splits

• data/: in-distribution (CCNEWS)
• ood/: out-of-distribution (Yahoo Answers)