Datasets:

Heliosoph
/

Quora-Question-Pairs

license: other
license_name: quora-qp-release-terms
license_link: https://quoradata.quora.com/First-Quora-Dataset-Release-Question-Pairs
pretty_name: Quora Question Pairs (2017 canonical release)
task_categories:
  - sentence-similarity
  - text-classification
language:
  - en
size_categories:
  - 100K<n<1M
tags:
  - embeddings
  - sentence-similarity
  - paraphrase
  - duplicate-questions
  - quora
  - qqp
source_datasets:
  - original

Quora Question Pairs — canonical 2017 release

A verbatim mirror of Quora's January 2017 Question Pairs release, packaged as a single tab-delimited file. No rows added, removed, or reordered relative to the upstream quora_duplicate_questions.tsv — only the hosting moved.

Re-hosted under Heliosoph for ingestion-pipeline stability — Quora's original CDN at qim.fs.quoracdn.net has been intermittently unreachable since the Kaggle competition wrapped, and the file has no checksumed permanent URL upstream.

Credit: Shankar Iyer, Nikhil Dandekar, and Kornél Csernai (Quora, 2017) — questions authored by Quora users.

Why a mirror?

The original release URL (qim.fs.quoracdn.net/quora_duplicate_questions.tsv) is the single canonical source, but it has gone offline for weeks at a time in past years, and there's no published checksum to verify a recovered copy against. Pinning a HuggingFace revision gives ingestion pipelines a stable, reproducible handle: the same SHA fetches the same bytes forever.

The file itself is unmodified — md5 / sha256 against the upstream copy (when it's online) will match exactly.

What this repo contains

quora_duplicate_questions.tsv     # 404,290 rows, ~55 MB, UTF-8, tab-delimited

One file, six columns, header row included:

Column	Type	Meaning
`id`	int	Row index (0-based).
`qid1`	int	Stable id of `question1` in Quora's internal question table.
`qid2`	int	Stable id of `question2`.
`question1`	string	First question, original casing + punctuation.
`question2`	string	Second question, original casing + punctuation.
`is_duplicate`	int (0/1)	Human label: 1 if the pair asks the same thing, else 0.

The qid1 / qid2 columns let you build a question-level view by deduping — the corpus is ~537k distinct questions arranged into 404k pairs.

How to use

Load with the datasets library:

from datasets import load_dataset

ds = load_dataset("Heliosoph/Quora-Question-Pairs", split="train")
print(ds[0])
# {'id': 0, 'qid1': 1, 'qid2': 2,
#  'question1': 'What is the step by step guide to invest in share market in india?',
#  'question2': 'What is the step by step guide to invest in share market?',
#  'is_duplicate': 0}

Or read the TSV directly with pandas:

import pandas as pd

df = pd.read_csv("quora_duplicate_questions.tsv", sep="\t")
print(df.shape)  # (404290, 6)
print(df["is_duplicate"].mean())  # ~0.369 — about 37% of pairs are duplicates

Dataset specs

	Spec
Rows	404,290 question pairs
Distinct questions	~537,000
Duplicate share	~36.9% labelled `is_duplicate = 1`
Encoding	UTF-8
Delimiter	Tab (`\t`)
Quoting	Minimal — embedded tabs in question text are rare but present; standard CSV/TSV readers handle them with double-quote escaping
File size	~55 MB on disk
Language	English

When to pick Quora Question Pairs

Sentence-embedding evaluation: encode question1 + question2, take cosine similarity of the L2-normalised vectors, threshold for a binary classifier, report accuracy / F1 against is_duplicate. This is the canonical paired-cosine evaluation that every English sentence embedder (MiniLM, BGE, Jina, E5, GTE) reports.
Duplicate-detection training: 404k labelled pairs is enough to fine-tune a small encoder end-to-end on a single GPU in hours.
Paraphrase identification: the label space (same question or not) is a clean two-class proxy for paraphrase / non-paraphrase.

For larger-scale retrieval (millions of passages) reach for MS MARCO instead — Quora QP is purpose-built for pairwise similarity, not document retrieval.

Label notes

The is_duplicate flag is human-annotated but not gold-standard: Quora's release post explicitly notes that some labels are noisy ("the ground-truth labels contain some amount of noise: they are not guaranteed to be perfect"). For benchmark numbers, report on the full corpus rather than a hand-cleaned subset to keep results comparable.

License

Non-standard research-use terms from Quora's 2017 release. Allows research, experimentation, and unmodified redistribution with attribution back to Quora. Commercial deployment is not explicitly addressed in the original announcement and should be reviewed separately.

The HuggingFace metadata tags this other because no SPDX identifier matches; the canonical terms are the release announcement.

Release announcement: First Quora Dataset Release: Question Pairs
Kaggle competition (same data): Quora Question Pairs