mehran-sarmadi's picture
Update README.md
23a604e verified
metadata
configs:
  - config_name: default
    data_files:
      - split: test
        path: qrels/test.jsonl
  - config_name: corpus
    data_files:
      - split: corpus
        path: corpus.jsonl
  - config_name: queries
    data_files:
      - split: queries
        path: queries.jsonl

Dataset Summary

CQADupstack-webmasters-Fa is a Persian (Farsi) dataset created for the Retrieval task, focusing on identifying duplicate or semantically similar questions within community question-answering (CQA) platforms. It is a translated version of the Webmasters StackExchange data from the English CQADupstack dataset and is part of the FaMTEB (Farsi Massive Text Embedding Benchmark).

  • Language(s): Persian (Farsi)
  • Task(s): Retrieval (Duplicate Question Retrieval)
  • Source: Translated from CQADupstack-Webmasters (BEIR benchmark) using Google Translate
  • Part of FaMTEB: Yes — as part of the BEIR-Fa collection

Supported Tasks and Leaderboards

The dataset is designed to test text embedding models' performance in retrieving duplicate or semantically equivalent questions in a technical domain (SEO, webmastering, site performance). It is benchmarked on the Persian MTEB Leaderboard (language: Persian).

Construction

This dataset was constructed via:

  • Extracting data from the Webmasters subforum of StackExchange (from the English CQADupstack dataset)
  • Translating the data into Persian using the Google Translate API
  • Retaining the original query-relevant pairs for Retrieval evaluation

As discussed in the FaMTEB paper, the entire BEIR-Fa collection (including this dataset) was evaluated using:

  • BM25 retrieval score comparison
  • GEMBA-DA framework leveraging LLMs to validate translation quality

These assessments indicate good fidelity in Persian translations.

Data Splits

The full CQADupstack-Fa collection has the following evaluation split:

  • Train: 0 samples
  • Dev: 0 samples
  • Test: 480,902 samples (across all domains)

The Webmasters-specific subset contains approximately 19.3k examples, though individual splits are not separately provided in the FaMTEB paper. For detailed splits, consult the dataset provider or Hugging Face dataset card.