WASIL: In-the-Wild Arabic Spoken Interactions with LLMs

The WASIL dataset has been released, providing a comprehensive collection of 8,529 Arabic spoken interaction prompts, including audio, ASR hypotheses, assistant responses, and user feedback, with a notable 14.2% dislike rate. It features a 2,000-turn test set that encompasses Modern Standard Arabic and four major dialects, with annotations for answerability and gold transcripts generated through multi-ASR agreement-guided post-editing. This resource is significant for practitioners as it allows for better evaluation of ASR systems and LLM responses, facilitating improvements in handling spoken Arabic interactions.