Dataset Summary
FarSick STS is a Persian (Farsi) dataset designed for the Semantic Textual Similarity (STS) task. It is a part of the FaMTEB (Farsi Massive Text Embedding Benchmark). The dataset was developed by translating and adapting the English SICK (Sentences Involving Compositional Knowledge) dataset, and it features Persian sentence pairs annotated for their degree of semantic relatedness.
- Language(s): Persian (Farsi)
- Task(s): Semantic Textual Similarity (STS)
- Source: Translated and adapted from the English SICK dataset
- Original Citation: FarSick dataset by Amin et al. (2022)
- Part of FaMTEB: Yes
Supported Tasks and Leaderboards
This dataset is primarily used to benchmark the ability of text embedding models to measure semantic similarity between Persian sentence pairs—especially in cases requiring compositional reasoning. Performance is benchmarked on the Persian MTEB Leaderboard on Hugging Face Spaces.
Construction
The construction process involved:
- Translating sentence pairs from the original English SICK dataset into Persian.
- Adapting the translations to ensure cultural and linguistic relevance.
- Preserving the focus on semantic compositionality, which includes reasoning about phrase structure and entailment.
- Annotating each Persian sentence pair with a similarity score reflecting their degree of semantic relatedness.
As noted in the FaMTEB paper, this dataset helps evaluate semantic understanding in Persian NLP models under compositional semantics conditions.
Data Splits
As defined in the FaMTEB paper (Table 5):
- Train: 0 samples
- Development (Dev): 0 samples
- Test: 8,566 samples