The table `raw.bedtime_screentime_sleep_debt` has 22 columns and 8,500 rows. Types and ranges come from the column metadata I pulled. The meanings are my reading of the column names, since the dataset has no data dictionary. No column has nulls. **Identity and demographics** - `user_id` (VARCHAR): identifier for each person, ranging from USR-00001 to USR-08500. - `age` (BIGINT): age in years, 18 to 65. - `gender` (VARCHAR): gender category with 3 values, including Female and Non-Binary. - `occupation_type` (VARCHAR): work or life situation with 5 values, such as "Corporate 9-to-5" and "Student". - `chronotype` (VARCHAR): natural sleep-timing preference with 3 values, such as Intermediate and Night Owl. **Bedtime phone and screen behaviour** - `bedtime_phone_minutes` (BIGINT): minutes of phone use around bedtime, 1 to 180. The card treats it as the temporal axis and flags it as person-identifying, so it should only be aggregated. - `primary_bedtime_app` (VARCHAR): the main app used at bedtime, with 6 values such as Instagram / Reddit and YouTube. - `screen_brightness_pct` (BIGINT): screen brightness as a percentage, 10 to 100. - `blue_light_filter_active` (BIGINT): a 0/1 flag for whether a blue-light filter was on. **Other daily habits** - `caffeine_post_5pm_mg` (BIGINT): caffeine consumed after 5pm, in mg, 0 to 250. - `physical_activity_min` (BIGINT): minutes of physical activity, 0 to 112. **Sleep outcomes** - `sleep_latency_min` (DOUBLE): minutes taken to fall asleep, 6.0 to 123.3. - `total_sleep_hours` (DOUBLE): total hours slept, 3.2 to 9.8. - `deep_sleep_pct` (DOUBLE): percentage of sleep spent in deep sleep, 8.1 to 28.0. - `rem_sleep_pct` (DOUBLE): percentage of sleep spent in REM sleep, 9.6 to 27.0. - `morning_alarm_snoozes` (BIGINT): number of alarm snoozes, 0 to 7. - `next_day_fatigue_score` (DOUBLE): fatigue score for the following day, 1.0 to 10.0. - `sleep_debt_category` (VARCHAR): a label with 4 values, from "Mild Deficit" up to "Severe Sleep Debt". It is derived from `total_sleep_hours`, so reuse it rather than recomputing it. **Ingestion metadata** (added when the data was loaded, not part of the observations) - `_ingestion_timestamp` (TIMESTAMP): when the data was loaded. It has a single value for the whole table, 2026-10-01 03:08:19. - `_batch_id` (VARCHAR): load batch identifier. It has a single value, `batch_86c069e6eb91`. - `_source_file` (VARCHAR): path of the source CSV file. It has a single value. - `_source_system` (VARCHAR): the system the data came from, which is "api" for every row. **Cautions** - The profiler labels some numeric columns as "categorical" or "text". That is a profiling quirk and does not change their stored numeric types. - The profiler labels `age` with a "duration" unit, but it is plainly age in years. - `user_id` shows about 7,297 distinct values against 8,500 rows. That figure is an approximate distinct count, so check uniqueness exactly before treating it as a strict key.