README.md

metadata

license: mit
dataset_info:
  features:
    - name: anchor
      dtype: string
    - name: positive
      dtype: string
    - name: negative
      dtype: string
    - name: sim_pos
      dtype: float64
    - name: sim_neg
      dtype: float64
    - name: len_anc
      dtype: int64
    - name: len_pos
      dtype: int64
    - name: len_neg
      dtype: int64
  splits:
    - name: train
      num_bytes: 614206347
      num_examples: 1000000
  download_size: 308842392
  dataset_size: 614206347
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Arabic 1 Million Triplets (curated):

This is a curated dataset to use in Arabic ColBERT and SBERT models (among other uses). In addition to anchor, positive and negative columns, the dataset has two columns: sim_pos and sim_neg which are cosine similarities between the anchor (query) and bothe positive and negative examples.
The last 3 columns are lengths (words) for each of the anchor, positive and negative examples. Length uses simple split on space, not tokens.

The cosine similarity uses an embedding model by AbderrahmanSkiredj1/Arabic_text_embedding_for_sts
(inspired by Omar Nicar) who made the first Arabic SBERT embeddings model and a triplets dataset based on NLI.

Why another dataset?

While training an Arabic ColBERT model using a sample from the mMARCO dataset, I noticed retrieval issues. It is true all these triplet datasets are translated, but quality was not up to expectation. I took the dataset used by the embedding model (which is NLI plus some 300K) and 1 million samples from mMARCO and removed lines that had seperate latin words/phrases and sampled 1 million rows of the combined data. Then I added the similiarity columns and lengths.
This should enable researchers and users to filter based on several criteria (including hard negatives). This is not saying the model used in similarities was perfect. In some cases, exmples annotated as negative were identical to the anchor/query. Adding the similarities columns took more time than training models.

Arabic SBERT and ColBERT models:

Filtered subsets based on certain criteria show impressive perfrmance. Models will be uploaded and linked from here when ready. If you saw earlier versions of triplets datasets under this account, they have been removed in favor of this one. If you downloaded or duplicated a triplets dataset from this account prior to Satuday 3 PM Jerusalem time on July 27th, 2024, you are also advised to get the updated version.