metadata

dataset_info:
  features:
    - name: anchor
      dtype: string
    - name: positive
      dtype: string
    - name: negative
      dtype: string
    - name: source
      dtype: string
    - name: anchor_len
      dtype: int64
    - name: positive_len
      dtype: int64
    - name: negative_len
      dtype: int64
  splits:
    - name: s1_akhooli
      num_bytes: 280399441
      num_examples: 1013295
    - name: s2_ArabicQuoraDuplicates
      num_bytes: 253139195
      num_examples: 850914
    - name: s3_wikimatrix
      num_bytes: 111840717
      num_examples: 222739
    - name: s4_tedtalks
      num_bytes: 64756065
      num_examples: 192699
    - name: s5_arqna
      num_bytes: 414087
      num_examples: 672
  download_size: 347683123
  dataset_size: 710549505
configs:
  - config_name: default
    data_files:
      - split: s1_akhooli
        path: data/s1_akhooli-*
      - split: s2_ArabicQuoraDuplicates
        path: data/s2_ArabicQuoraDuplicates-*
      - split: s3_wikimatrix
        path: data/s3_wikimatrix-*
      - split: s4_tedtalks
        path: data/s4_tedtalks-*
      - split: s5_arqna
        path: data/s5_arqna-*

SILMA Arabic Triplets Dataset - `v1.0`

Overview

The SILMA Arabic Triplets Dataset - v1.0 is a high-quality, diverse dataset specifically curated for training and training embedding models for semantic search tasks in the Arabic language.

The dataset contains more than 2.25M records (2,280,319 records).

This dataset includes triplets in the form of anchor, positive, and negative samples, designed to enhance models in learning semantic similarity and dissimilarity.

The dataset consists of five unique splits sourced from diverse domains, providing a robust foundation for building Arabic-language models that require nuanced understanding of text relationships.

Dataset Structure

The dataset is structured as a DatasetDict containing five different sources, each with their own distinct rows and domains:

Splits

Here are descriptions with associated representative examples from each split of the dataset:

Split 1: Akhooli

Description: A curated set of the original akhooli associated with a defined negative sample for each record.

Size: 1,013,295 samples

Sample

{
  "anchor": "صبي صغير وفتاة صغيرة يركبان دراجتيهما على الرصيف مع عجلات مساعدة.",
  "positive": "فتى وفتاة يتعلمون ركوب الدراجات",
  "negative": "الصبي الصغير يصل إلى العصا من الفتاة وهو يدير سباق التتابع"
}

Split 2: ArabicQuoraDuplicates

Description: A curated set of the original ArabicQuoraDuplicates source after cleaning and filtering out some outliers.

Size: 850,914 samples

Sample

{
  "anchor": "ثلاثة رجال يلعبون كرة السلة",
  "positive": "ثلاثة رجال يلعبون لعبة كرة السلة",
  "negative": "رجلين يرتديان ملابس غريبة يقفزان على ملعب كرة السلة"
}

Split 3: WikiMatrix

Description: A curated set of the original WikiMatrix associated with a defined negative sample for each record.

Size: 222,739 samples

Sample

{
  "anchor": "When multi-party elections began in the 1940s, the share of women in the legislature fell, and the 4% share of parliamentary seats gained in 1935 was not reached again until 1999.",
  "positive": "وعندما بدأت انتخابات الأحزاب المتعددة في الأربعينيات من القرن الماضي، انخفضت حصة المرأة في المجلس التشريعي، ولم يتم التوصل إلى نسبة 4 في المائة من المقاعد البرلمانية التي اكتُسبت في عام 1935 مرة أخرى حتى عام 1999.",
  "negative": "الانتخابات النيابية والبلدية تجري كل أربع سنوات منذ عودة العمل بالدستور في عام 2002 عندما تم منح المرأة أيضا التصويت لأول مرة كجزء من الإصلاحات الديمقراطية التي تعهد بها الملك حمد."
}

Split 4: TedTalks

Description: A curated set of the original TedTalks2020-en:ar associated with a defined negative sample for each record.

Size: 192,699 samples

Sample

{
  "anchor": "But if we think about it, we are actually recently arrived guests on this planet, the human species.",
  "positive": "ولكن إذا أمعنا النظر فيها، فسنجد أننا قد وصلنا إلى مرحلة أننا ضيوف على هذا الكوكب، أقصد بذلك الجنس البشري.",
  "negative": "ولكن إذا أردنا أن نستكشف نهايات حدود هذا الكوكب، علينا أن نعيش هناك."
}

Split 5: QnA

Description: A small QnA Arabic triplets dataset.

Size: 672 samples

Sample

{
  "anchor": "ما هي منظمة غرب إفريقيا التي تنتمي إليها غينيا كوناكري؟",
  "positive": "غينيا كوناكري عضو في الأمم المتحدة والجماعة الاقتصادية لدول غرب إفريقيا.",
  "negative": "أي بلد يبعد حوالي 4200 كم عن غينيا كوناكري؟"
}

Column Descriptions

Each dataset in the SILMA Arabic Triplets Dataset - v1.0 contains the following columns:

`anchor`

Description: The main reference sentence or query in the triplet. It serves as the anchor point to which the positive and negative samples are compared.
Type: String

`positive`

Description: A sentence that is semantically similar to the anchor sentence. This sentence is the "positive" counterpart that shares meaning or context with the anchor.
Type: String

`negative`

Description: A sentence that is semantically dissimilar to the anchor sentence. It provides contrast by being unrelated in meaning to the anchor.
Type: String

`source`

Description: The source or dataset from which the triplet is derived. This helps identify the domain or origin of the triplet data.
Type: String

`anchor_len`

Description: The number of words or tokens in the anchor sentence. This helps in analyzing sentence length and complexity across the dataset.
Type: Integer

`positive_len`

Description: The number of words or tokens in the positive sentence. This gives insights into the sentence length of semantically similar examples.
Type: Integer

`negative_len`

Description: The number of words or tokens in the negative sentence. This column reflects the length of the dissimilar sentence in the triplet.
Type: Integer

Summary

Each triplet consists of an anchor sentence, a positive sentence that is semantically similar, and a negative sentence that is semantically dissimilar. The length of each sentence is captured in anchor_len, positive_len, and negative_len, and the source column identifies the origin of the triplet, providing additional context for the data.

Use Cases

This dataset is ideal for:

Training embeddings for semantic search systems.
Fine-tuning language models on Arabic textual similarities.
Evaluating embedding-based retrieval models, particularly in triplet-based tasks.

The dataset spans different domains, such as general knowledge (WikiMatrix), public speaking (TedTalks), and QA forums (ArabicQuoraDuplicates), ensuring a diverse understanding of Arabic language semantics.