---
configs:
  - config_name: default
    data_files:
      - split: train
        path: qrels/train.jsonl
      - split: dev
        path: qrels/dev.jsonl
      - 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

**FEVER-Fa** is a Persian (Farsi) dataset designed for the **Retrieval** task. It is a key component of the [FaMTEB (Farsi Massive Text Embedding Benchmark)](https://huggingface.co/spaces/mteb/leaderboard) and represents a translated version of the original English FEVER dataset. This dataset is specifically tailored for evaluating models on automatic fact-checking by requiring the retrieval of evidential sentences from a pre-processed Wikipedia corpus that support or refute given claims.

* **Language(s):** Persian (Farsi)
* **Task(s):** Retrieval (Fact Checking, Evidence Retrieval)
* **Source:** Translated from the English [FEVER dataset](https://fever.ai/) using Google Translate.
* **Part of FaMTEB:** Yes (specifically, part of the BEIR-Fa collection within FaMTEB)

## Supported Tasks and Leaderboards

This dataset is primarily used to evaluate the performance of text embedding models on the **Retrieval** task. Model performance can be benchmarked and compared on the [Persian MTEB Leaderboard on Hugging Face Spaces](https://huggingface.co/spaces/mteb/leaderboard) (filter by language: Persian).

## Construction

The **FEVER-Fa** dataset was created by machine-translating the original English FEVER (Fact Extraction and VERification) dataset into Persian. The translation was performed using the Google Translate API.

As detailed in the "FaMTEB: Massive Text Embedding Benchmark in Persian Language" paper, the quality of the BEIR-Fa collection (of which FEVER-Fa is a part) underwent rigorous evaluation. This included:
1.  Comparing BM25 retrieval scores between the original English versions and the translated Persian versions, which showed comparable performance.
2.  Utilizing Large Language Models (LLMs) for a direct assessment of translation quality (GEMBA-DA framework), which indicated good overall translation quality, competitive with translations produced by other prominent LLMs.

## Data Splits

The data is split into training and test sets as defined in the FaMTEB paper (Table 5):

* **Train:** 5,556,643 samples
* **Development (Dev):** 0 samples
* **Test:** 5,424,495 samples