XStoryCloze-FTP: First-Token-Prediction Story Completion for Language Models

A reframing of the XStoryCloze dataset for first-token-prediction (FTP) tasks. Each example is formatted as a single text sequence ending with 'Answer：' to allow a model to predict the answer as one token. The dataset was created by mikaberidze and was last updated on 2026-05-18.

Use Cases

• Benchmarking language model performance on story cloze tasks based on the FTP reframing.
• Training models for next-token prediction on structured multiple-choice sequences based on the described format.
• Evaluating model robustness on cross-lingual narrative understanding based on the dataset's multilingual nature.
• Fine-tuning models for specific answer token prediction based on the 'Answer：' prompt structure.

Strengths

• Format matches the lm-evaluation-harness's MCQA template byte-for-byte, ensuring compatibility with standard evaluation frameworks.
• Examples are specifically formatted for first-token-prediction, a targeted evaluation scenario.
• Dataset was last updated on 2026-05-18 12:15:39.

Limitations

• Row count is unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Reframing of the existing juletxara/xstory_cloze dataset.

Freshness

Last updated 2026-05-18 12:15:39.