sentence-transformers/quora-duplicates-mining

Dataset Card for Quora Duplicate Questions

This dataset contains the Quora Question Pairs dataset in a format that is easily used with the ParaphraseMiningEvaluator evaluator in Sentence Transformers. The data was originally created by Quora for this Kaggle Competition.

Usage

from datasets import load_dataset
from sentence_transformers.SentenceTransformer import SentenceTransformer
from sentence_transformers.evaluation import ParaphraseMiningEvaluator

# Load the Quora Duplicates Mining dataset
questions_dataset = load_dataset("sentence-transformers/quora-duplicates-mining", "questions", split="dev")
duplicates_dataset = load_dataset("sentence-transformers/quora-duplicates-mining", "duplicates", split="dev")

# Create a mapping from qid to question & a list of duplicates (qid1, qid2)
qid_to_questions = dict(zip(questions_dataset["qid"], questions_dataset["question"]))
duplicates = list(zip(duplicates_dataset["qid1"], duplicates_dataset["qid2"]))

# Initialize the paraphrase mining evaluator
paraphrase_mining_evaluator = ParaphraseMiningEvaluator(qid_to_questions, duplicates, name="quora-duplicates-dev")

# Load a model to evaluate
model = SentenceTransformer("all-MiniLM-L6-v2")
results = paraphrase_mining_evaluator(model)
print(results)

{
  'quora-duplicates-dev_average_precision': 0.5537837023752262,
  'quora-duplicates-dev_f1': 0.542585123346778,
  'quora-duplicates-dev_precision': 0.5112918195076678,
  'quora-duplicates-dev_recall': 0.5779587350751861,
  'quora-duplicates-dev_threshold': 0.8290803134441376,
}

Dataset Subsets

questions subset

• Columns: "question", "qid"
• Column types: str, str
• Examples:

  {
    'question': 'How do I prepare for TCS IT Wiz?',
    'qid': '107646',
  }
  

• Collection strategy: A direct copy of the quora-IR-dataset/duplicate-mining as generated from create_splits.py.
• Deduplified: No

duplicates subset

• Columns: "qid1", "qid2"
• Column types: str, str
• Examples:

  {
    'qid1': '43345',
    'qid2': '43346',
  }
  

• Collection strategy: A direct copy of the quora-IR-dataset/duplicate-mining as generated from create_splits.py.
• Deduplified: No