dataset_info:
config_name: triplet
features:
- name: query
dtype: string
- name: answer
dtype: string
- name: negative
dtype: string
splits:
- name: train
num_bytes: 112942322
num_examples: 76709
download_size: 33017802
dataset_size: 112942322
configs:
- config_name: triplet
data_files:
- split: train
path: triplet/train-*
license: apache-2.0
Dataset Summary
This dataset is an enhanced version of the Arabic Natural Questions dataset, augmented with hard negatives mined using the Sentence Transformers library. Hard negatives are examples that are particularly challenging for models—they're similar to the correct answer but are incorrect, providing a stronger training signal. This dataset is particularly useful for improving model performance in tasks like question-answering, semantic similarity, and information retrieval. Supported Tasks and Leaderboards
Task: Semantic Textual Similarity, Information Retrieval, Question-Answering Leaderboards: This dataset does not currently have associated leaderboards.
Languages The dataset is in Arabic.
Dataset Structure
- query: The original question or query.
- answer: The correct answer associated with the query.
- negative: The mined hard negative, which is a text similar to the correct answer but incorrect.
Dataset Creation
The hard negatives were mined to provide a challenging set of examples for training models that need to differentiate between very similar texts. This approach strengthens the model’s ability to make fine-grained distinctions, which is essential in tasks such as question-answering and semantic search.
- Initial Dataset: The original dataset used is the Arabic Natural Questions dataset.
- Mining Process: Hard negatives were mined using the Sentence Transformers mine_hard_negatives utility with the following parameters: range_min: 10 range_max: 50 max_score: 0.8 margin: 0.1 sampling_strategy: random
Annotations
No additional human annotation was performed; the hard negatives were generated using Hard Negatives Mining utility from Sentence Transformer