Datasets:
| language: | |
| - fa | |
| license: apache-2.0 | |
| size_categories: | |
| - 100K<n<1M | |
| task_categories: | |
| - question-answering | |
| - text-classification | |
| # 📘 PerCoR: Persian Commonsense Reasoning (Multiple-Choice Sentence Completion) | |
| Paper: [PerCoR: Evaluating Commonsense Reasoning in Persian via Multiple-Choice Sentence Completion](https://huggingface.co/papers/2510.22616) | |
| **PerCoR** is a large-scale Persian benchmark for **commonsense reasoning** in a **4-choice sentence-completion** format. | |
| It contains **~106K** examples from **40+** Persian websites across news, culture, lifestyle, tech, religion, travel, and more. | |
| Each instance provides a **prefix** (context) and **four candidate completions** — one correct and three distractors. | |
| --- | |
| ## 📦 What’s inside | |
| - 🧮 **Total size:** ~106K multiple-choice instances | |
| - 📊 **Splits:** `train` 86,217 • `validation` 10,000 • `test` 10,000 | |
| - 🧱 **Format:** single passage/prefix + 4 completions (A–D / 0–3) with one correct answer | |
| - 🧠 **Human accuracy:** ~89% on a random subset | |
| > 💡 *The dataset is designed to be difficult for LLMs while remaining answerable by humans; no LLM text is used to generate distractors (reducing generation-style biases).* |