A large, consolidated collection of English sentence pairs for training and evaluating semantic similarity, retrieval, and re-ranking models. It merges widely used benchmarks into a single schema with consistent fields and ready-made splits.

Dataset Details

Dataset Description

Name: langcache-sentencepairs-v1
Summary: Sentence-pair dataset created to fine-tune encoder-based embedding and re-ranking models. It combines multiple high-quality corpora spanning diverse styles (short questions, long paraphrases, Twitter, adversarial pairs, technical queries, news headlines, etc.), with both positive and negative examples and preserved splits.
Curated by: Redis
Shared by: Aditeya Baral
Language(s): English
License: Apache-2.0
Homepage / Repository: https://huggingface.co/datasets/redis/langcache-sentencepairs-v1

Configs and coverage

all: Unified view over all sources with extra metadata columns (id, source, source_idx).
Source-specific configs: apt, mrpc, parade, paws, pit2015, qqp, sick, stsb.

Size & splits (overall)
Total ~1.12M pairs: ~1.05M train, 8.4k validation, 62k test. See per-config sizes in the viewer.

Dataset Sources

APT (Adversarial Paraphrasing Task) — Paper | Dataset
MRPC (Microsoft Research Paraphrase Corpus) — Paper | Dataset
PARADE (Paraphrase Identification requiring Domain Knowledge) — Paper | Dataset
PAWS (Paraphrase Adversaries from Word Scrambling) — Paper | Dataset
PIT2015 (SemEval 2015 Twitter Paraphrase) — Website | Dataset
QQP (Quora Question Pairs) — Website | Dataset
SICK (Sentences Involving Compositional Knowledge) — Website | Dataset
STS-B (Semantic Textual Similarity Benchmark) — Website | Dataset

Uses

Train/fine-tune sentence encoders for semantic retrieval and re-ranking.
Supervised sentence-pair classification tasks like paraphrase detection.
Evaluation of semantic similarity and building general-purpose retrieval and ranking systems.

Direct Use

from datasets import load_dataset

# Unified corpus
ds = load_dataset("redis/langcache-sentencepairs-v1", "all")

# A single source, e.g., PAWS
paws = load_dataset("redis/langcache-sentencepairs-v1", "paws")

# Columns: sentence1, sentence2, label (+ idx, source_idx in 'all')

Out-of-Scope Use

Non-English or multilingual modeling: The dataset is entirely in English and will not perform well for training or evaluating multilingual models.
Uncalibrated similarity regression: The STS-B portion has been integerized in this release, so it should not be used for fine-grained regression tasks requiring the original continuous similarity scores.

Dataset Structure

Fields

sentence1 (string) — First sentence.
sentence2 (string) — Second sentence.
label (int64) — Task label. 1 ≈ paraphrase/similar, 0 ≈ non-paraphrase/dissimilar. For sources with continuous similarity (e.g., STS-B), labels are integerized in this release; consult the source subset if you need original continuous scores.
(config all only):
- id (string) — Dataset identifier. Follows the pattern langcache_{split}_{row number}.
- source (string) — Source dataset name.
- source_idx (int64) — Source-local row id.

Splits

train, validation (where available), test — original dataset splits preserved whenever provided by the source.

Schemas by config

all: 5 columns (idx, source_idx, sentence1, sentence2, label).
All other configs: 3 columns (sentence1, sentence2, label).

Dataset Creation

Curation Rationale

To fine-tune stronger encoder models for retrieval and re-ranking, we curated a large, diverse pool of labeled sentence pairs (positives & negatives) covering multiple real-world styles and domains. Consolidating canonical benchmarks into a single schema reduces engineering overhead and encourages generalization beyond any single dataset.

Source Data

Data Collection and Processing

Ingested each selected dataset and preserved original splits when available.
Normalized to a common schema; no manual relabeling was performed.
Merged into all with added source and source_idx for traceability.

Who are the source data producers?

Original creators of the upstream datasets (e.g., Microsoft Research for MRPC, Quora for QQP, Google Research for PAWS, etc.).

Personal and Sensitive Information

The corpus may include public-text sentences that mention people, organizations, or places (e.g., news, Wikipedia, tweets). It is not intended for identifying or inferring sensitive attributes of individuals. If you require strict PII controls, filter or exclude sources accordingly before downstream use.

Bias, Risks, and Limitations

Label noise: Some sources include noisily labeled pairs (e.g., PAWS large weakly-labeled set).
Granularity mismatch: STS-B's continuous similarity is represented as integers here; treat with care if you need fine-grained scoring.
English-only: Not suitable for multilingual evaluation without adaptation.

Recommendations

Use the all configuration for large-scale training, but be aware that some datasets dominate in size (e.g., PAWS, QQP). Apply sampling or weighting if you want balanced learning across domains.
Treat STS-B labels with caution: they are integerized in this release. For regression-style similarity scoring, use the original STS-B dataset.
This dataset is best suited for training retrieval and re-ranking models. Avoid re-purposing it for unrelated tasks (e.g., user profiling, sensitive attribute prediction, or multilingual training).
Track the source field (in the all config) during training to analyze how performance varies by dataset type, which can guide fine-tuning or domain adaptation.

Citation

If you use this dataset, please cite the Hugging Face entry and the original upstream datasets you rely on.

BibTeX:

@misc{langcache_sentencepairs_v1_2025,
  title        = {langcache-sentencepairs-v1},
  author       = {Baral, Aditeya and Redis},
  howpublished = {\url{https://huggingface.co/datasets/redis/langcache-sentencepairs-v1}},
  year         = {2025},
  note         = {Version 1}
}