---
    pretty_name: grail-interactions
    tags:
    - grail
    - recommendation
    - slate
    - bandit
    - human-preference
    license: other
    language:
    - en
    task_categories:
    - recommendation
    task_ids:
    - slate-recommendation
    size_categories:
    - 10K<n<100K
    ---

    # grail-interactions

    Hugging Face dataset built from platform interactions with a **70/30 train/validation** split.

    ## Split sizes
    {
  "train": 41240,
  "validation": 4578
}

    ## What each row contains
    - **Profile one-liner** (human-readable demographics)  
    - **Current item** (title + id)  
    - **History up to current** (with watch seconds & total length)  
    - **Options** (slate shown at this step; names preferred, ids as fallback)

    ## Key columns (produced by the builder)
    - `current_video_id`, `current_video_title`  
    - `watched_vids_json`, `watched_detailed_json`  
    - `slate_items_json`, `slate_ids_json`, `slate_titles_json`, `slate_text`, `n_options`  
    - `viewer_profile_sentence`  
    - `state_text` (PROFILE + CURRENTLY WATCHING + HISTORY + OPTIONS)

    ## Example columns (first 30)
    ['worker_id', 'response_id', 'assignment_id', 'hit_id', 'qid', 'ytid', 'q87', 'q89', 'q5_2', 'q5_5', 'q5_11', 'q5_12', 'q58_1', 'q58_2', 'q58_3', 'q58_4', 'q89_1', 'q90', 'q6', 'q6_46_text', 'q56', 'q77', 'q8', 'q78', 'q9', 'q10', 'q12_1', 'q12_2', 'q12_3', 'q13']