UzFact: A Benchmark Dataset for Fact Extraction and Verification (FEVER) in the Uzbek Language
Description
The Knowledge Base was compiled by selecting 542 encyclopedic articles, yielding 2,902 uniquely indexed sentences. A balanced corpus of 1,500 claims was authored. Each verifiable claim is cross-referenced with exact document IDs and sentence line indices. Quality assurance was performed independently by five annotators and reviewed by native Uzbek linguists, yielding Fleiss' κ = 0.7121 All dataset files are serialized in **JSON Lines (`.jsonl`)** format, where each line corresponds to an independent JSON object. ### A. Claims File (`uzfact_standard_1500.jsonl` / Splits) * **`id`** (`int`): Unique sample identifier (range: 1 to 1500). * **`claim`** (`str`): Natural language assertion to be verified. * **`label`** (`str`): Verification verdict across three balanced classes: * `SUPPORTS` — The claim is substantiated by explicit evidence in the Knowledge Base (500 samples). * `REFUTES` — The claim contradicts the factual evidence in the Knowledge Base (500 samples). * `NOT ENOUGH INFO` — The Knowledge Base lacks sufficient evidence to either verify or refute the claim (500 samples). * **`evidence`** (`list`): Hierarchical evidence pointer coordinates following the official FEVER schema: * Format: `[[[annotation_id, evidence_id, "Document_ID", sentence_index]]]` * `Document_ID`: Canonical article identifier matching the `id` field in `uzfact_kb_1500.jsonl`. * `sentence_index`: Zero-based line index of the evidentiary sentence within the source document. * *Note:* For `NOT ENOUGH INFO` assertions, the `evidence` field is strictly an empty list (`[]`).