CLOTH contains 99,433 cloze questions sourced from middle and high school English exams. The dataset is split into training, validation, and test sets, with 76,850, 11,067, and 11,516 questions respectively. It was created by AndyChiang and is intended for reading comprehension tasks.
Use Cases
- Train reading comprehension models on 99,433 cloze questions to predict missing words in exam text.
- Analyze the difficulty distribution of cloze questions between the middle school and high school subsets.
- Benchmark language model performance on the separate validation (11,067 questions) and test (11,516 questions) splits.
Strengths
- Large scale with 99,433 total cloze questions for model training and evaluation.
- Clear data split with 76,850 training, 11,067 validation, and 11,516 test questions.
- Structured categorization into 22,056 middle school and 54,794 high school questions for targeted analysis.
Limitations
- No column or feature descriptions are provided, limiting understanding of the data structure.
- The source and collection methodology for the exam questions are not detailed, raising provenance questions.
- The dataset is static with no stated update frequency, potentially becoming stale for evolving language models.
Provenance
- Source
- https://www.cs.cmu.edu/~glai1/data/cloth/
- Collection Method
- Collection of cloze questions from middle and high school English exams.
- Time Range
- null
- Freshness
- null
- Geography
- null