Dataset documentation

Field descriptions

A complete reference for KuaiLive-M3's shared entities, live-streaming records, short-video behaviors, metadata, and multimodal embeddings.

18tables & collections

100documented fields

2content domains

Privacy

All user, content, creator, room, and segment identifiers are anonymized. Entity IDs are remapped to consecutive integers starting from 1 unless otherwise noted.

← Back to overview Open dataset ↗

18 of 18 tables shown

Shared

author_profile.csv

Profile attributes for creators appearing in either content domain.

6fields

Field	Type	Description
`author_id`	int	Anonymized creator ID.
`is_photo_author`	int	Whether the creator publishes short videos (1=yes, 0=no).
`is_live_author`	int	Whether the creator hosts live streams (1=yes, 0=no).
`gender`	str	Gender: M, F, or U (unknown).
`age_segment`	str	Age interval: 0-12, 12-17, 18-23, 24-30, 31-40, 41-49, 50+, or unknown.
`fans_user_num`	str	Binned follower count, from 0-1k to 1000w+.

Shared

user_id_set.csv

The anonymized user population shared across both domains.

1fields

NoteApproximately 22K users, deterministically sampled by hash from a larger user pool.

Field	Type	Description
`user_id`	int	Anonymized user ID.

Live

live_interaction.csv

One row represents an aggregated user viewing session in a live room.

20fields

Field	Type	Description
`user_id`	int	Anonymized user ID.
`author_id`	int	Anonymized streamer ID.
`live_id`	int	Anonymized live-room ID.
`live_play_start_timestamp`	timestamp	Viewing start time with millisecond precision.
`live_play_end_timestamp`	timestamp	Viewing end time with millisecond precision.
`p_date`	str	Date partition in yyyyMMdd format.
`live_source_category`	str	The page from which the user entered the room, e.g. follow, nearby, feed, search, profile, or live square.
`enter_live_action`	str	Entry action: click, slide up/down, re-enter, auto-enter, or unknown.
`is_auto_play`	int	Whether the room started through autoplay.
`is_follow_enter`	int	Whether the user followed the streamer upon entry.
`is_follow_leave`	int	Whether the user followed the streamer upon leaving.
`play_duration`	int	Watch duration in milliseconds.
`like_cnt`	int	Number of likes during the session.
`comment_cnt`	int	Number of comments, including voice comments.
`send_gift_cnt`	int	Number of gift-sending actions.
`send_gift_num`	int	Number of virtual gift items sent.
`follow_author_cnt`	int	Number of follow actions during the session.
`cancel_follow_author_cnt`	int	Number of unfollow actions during the session.
`share_cnt`	int	Number of sharing actions.
`report_live_cnt`	int	Number of room-reporting actions.

Live

live_show.parquet

Feed impressions where a user saw a live-room cover without entering the room; useful for negative sampling.

4fields

Field	Type	Description
`live_id`	int	Anonymized live-room ID.
`user_id`	int	Anonymized user ID.
`author_id`	int	Anonymized streamer ID.
`show_timestamp`	timestamp	Impression timestamp with millisecond precision.

Live

live_comment.csv

Timestamped text comments posted by users in live rooms.

5fields

Field	Type	Description
`live_stream_id`	int	Anonymized live-room ID.
`user_id`	int	Anonymized commenting user ID.
`author_id`	int	Anonymized streamer ID.
`content`	str	Comment text.
`comment_ts`	int	Unix timestamp in milliseconds.

Live

live_like.csv

Event-level likes with millisecond timestamps.

3fields

Field	Type	Description
`user_id`	int	Anonymized user ID.
`live_id`	int	Anonymized live-room ID.
`like_timestamp`	timestamp	Like timestamp with millisecond precision.

Live

live_share.csv

Event-level sharing actions within live rooms.

5fields

Field	Type	Description
`user_id`	int	Anonymized user ID.
`live_id`	int	Anonymized live-room ID.
`author_id`	int	Anonymized streamer ID.
`share_timestamp`	timestamp	Share timestamp with millisecond precision.
`is_share_success`	int	Whether the sharing action succeeded.

Live

live_questionnaire.csv

Explicit feedback collected from users while watching live streams.

5fields

NoteThe release includes questionnaire responses related to following and gifting preferences.

Field	Type	Description
`author_id`	int	Anonymized streamer ID.
`live_stream_id`	int	Anonymized live-room ID.
`user_id`	int	Anonymized respondent ID.
`select_option`	str	Answer to the first selection question.
`second_select_option`	str	Answer to the second selection question.

Live

live_room_set.csv

All live-room IDs covered by the dataset.

1fields

NoteApproximately 6.56 million live rooms.

Field	Type	Description
`live_id`	int	Anonymized live-room ID.

Live

live_room_meta.parquet

Room titles, lifecycle timestamps, streamer mapping, and broadcast duration.

6fields

Field	Type	Description
`live_id`	int	Anonymized live-room ID.
`author_id`	int	Anonymized streamer ID.
`live_name`	str	Live-room title; empty when unavailable.
`start_timestamp`	timestamp	Room start time with millisecond precision.
`end_timestamp`	timestamp	Room end time; NULL when the room had not ended.
`live_duration`	int	Broadcast duration in milliseconds; cross-day streams are partitioned by date.

Live

live_emb_64.parquet

Aggregated content representation for each live room.

2fields

Field	Type	Description
`live_id`	int	Anonymized live-room ID.
`embedding`	array[float]	64-dimensional room-level embedding.

Live

live_emb_128_ts/

Timestamped content representations for temporally evolving live-room segments.

4fields

NoteStored in 18 Parquet shards. Each timestamp denotes the end of the corresponding, variable-length segment. Only segments with at least one viewer are retained.

Field	Type	Description
`live_stream_id`	str	Anonymized live-room ID.
`segment_id`	str	Anonymized segment ID; each room contains multiple temporally ordered segments.
`feature_128`	array[float]	128-dimensional segment embedding obtained by PCA from an original 2,307-dimensional multimodal feature.
`timestamp`	int	Segment end timestamp in Unix milliseconds, for alignment with user viewing windows.

Short video

photo_interaction.csv

Daily aggregated interactions between a user and a short video; only valid plays are retained.

14fields

Field	Type	Description
`user_id`	int	Anonymized user ID.
`photo_id`	int	Anonymized short-video ID.
`author_id`	int	Anonymized creator ID.
`show_cnt`	int	Number of impressions.
`complete_play_cnt`	int	Number of complete plays.
`play_progress`	float	Mean playback progress in [0, 1].
`like_cnt`	int	Number of likes.
`cancel_like_cnt`	int	Number of like cancellations.
`direct_comment_cnt`	int	Number of direct comments.
`reply_comment_cnt`	int	Number of replies to other comments.
`comment_stay_duration`	int	Time spent in the comment area, in milliseconds.
`follow_cnt`	int	Number of creator follows attributed to the video.
`cancel_follow_cnt`	int	Number of creator unfollows attributed to the video.
`share_cnt`	int	Number of successful shares.

Short video

photo_play.parquet

Event-level short-video playback records; a user-video pair may appear multiple times.

10fields

NoteA common positive criterion is playback progress ≥10%, computed from the enter/leave timestamps and video duration.

Field	Type	Description
`user_id`	int	Anonymized user ID.
`photo_id`	int	Anonymized short-video ID.
`author_id`	int	Anonymized creator ID.
`enter_timestamp`	timestamp	Playback-page entry time with millisecond precision.
`leave_timestamp`	timestamp	Playback-page exit time with millisecond precision.
`enter_play_type`	str	Entry action: click, slide up/down, autoplay, re-enter, or unknown.
`is_complete_play`	int	Whether playback completed.
`like_status_type`	int	Like-state transition: new like, cancellation, never liked, or remained liked.
`is_follow_before_play`	int	Whether the user followed the creator before playback.
`is_follow_after_play`	int	Whether the user followed the creator after playback.

Short video

photo_id_set.csv

All short-video IDs covered by the dataset.

1fields

NoteApproximately 6.74 million short videos.

Field	Type	Description
`photo_id`	int	Anonymized short-video ID.

Short video

photo_meta.parquet

Video duration, creator mapping, display format, titles, OCR, and subtitle text.

6fields

Field	Type	Description
`photo_id`	int	Anonymized short-video ID.
`author_id`	int	Anonymized creator ID.
`duration`	int	Video duration in milliseconds.
`display_type`	int	Encoded display format.
`cover_title`	str	Cover-title text.
`video_texts`	str	Text observed in the video, including OCR and subtitles.

Short video

photo_tag.csv

A four-level hierarchical content taxonomy for short videos.

5fields

Field	Type	Description
`photo_id`	int	Anonymized short-video ID.
`first_level_category_name`	str	Level-one category name.
`second_level_category_name`	str	Level-two category name.
`third_level_category_name`	str	Level-three category name.
`fourth_level_category_name`	str	Level-four category name.

Short video

photo_emb_128.parquet

Video-level multimodal content representation.

2fields

Field	Type	Description
`photo_id`	int	Anonymized short-video ID.
`feature`	array[float]	128-dimensional short-video embedding.