File size: 2,102 Bytes
bb39db8 207cb6d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 | ---
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
**MSMARCO-Fa** is a Persian (Farsi) dataset created for the **Retrieval** task, particularly focusing on **web search** and **document ranking**. It is a translated version of the original English **MS MARCO (Microsoft MAchine Reading COmprehension)** dataset and is a key part of the [FaMTEB (Farsi Massive Text Embedding Benchmark)](https://huggingface.co/spaces/mteb/leaderboard), under the **BEIR-Fa** collection.
- **Language(s):** Persian (Farsi)
- **Task(s):** Retrieval (Web Search, Document Ranking)
- **Source:** Translated from the English MS MARCO dataset
- **Part of FaMTEB:** Yes — under BEIR-Fa
## Supported Tasks and Leaderboards
This dataset is used to evaluate the effectiveness of **text embedding models** in **ranking web documents** based on relevance to user queries, simulating real-world search engine applications. Benchmarking is available via the **Persian MTEB Leaderboard** (language: Persian).
## Construction
The dataset was built by:
- **Translating** the original English MS MARCO dataset to Persian using the **Google Translate API**
- Preserving original relevance annotations, where some passages are **human-judged** as relevant to each query
As described in the *FaMTEB* paper:
- Translation quality was evaluated by **BM25 retrieval score comparison** with the English dataset
- Further validation was done using **LLM-based assessments (GEMBA-DA framework)**
- This dataset is similar in structure to mMARCO, but focused solely on the **Persian language**
## Data Splits
Based on FaMTEB paper (Table 5):
- **Train:** 9,374,574 samples
- **Dev:** 0 samples
- **Test:** 8,845,925 samples
**Total:** ~9.9 million examples
|