public videos, comments, comment threads and danmaku from Bilibili, scraped continuously since October 2026. comments, danmaku and stats history are uploaded every day, videos and the stats snapshot weekly.

from datasets import load_dataset

danmaku = load_dataset("tiagozip/bilibili", "danmaku", split="train", streaming=True)
videos = load_dataset("tiagozip/bilibili", "videos", split="train", streaming=True)

format

config	notes
videos	one file per month of `pubdate`, `data/videos/YYYY-MM.parquet`
comments	split by scrape day, `data/comments/fetched=YYYY-MM-DD/*.parquet`
danmaku	split by scrape day, `data/danmaku/fetched=YYYY-MM-DD/*.parquet`
stats	snapshot, latest counters per video
stats_history	every counter reading, split by scrape day, `data/stats_history/fetched=YYYY-MM-DD/*.parquet`

comments and danmaku are split by the day they were scraped (fetched_at), not by when they were posted. the scraper revisits videos, so the same comment or danmaku can appear in more than one day's folder with newer counters. to get one row each, dedupe on (oid, type, rpid) for comments and (cid, dmid) for danmaku and keep the row with the newest fetched_at. inside the a-* and b-* files of a single day, a row can also repeat, and the same rule applies.

a video (aid) has one or more parts, and each part has its own cid. danmaku belong to a part, so join them to videos through videos.cids. comments on a video have type = 1 and oid = aid.

videos

column	type	notes
`aid`	uint64	numeric video id
`bvid`	string	the `BV...` id used in urls
`mid`	uint64	uploader id
`title`	string
`intro`	string	description
`tid`	uint32	category id
`tname`	string	category name
`copyright`	uint8	1 original, 2 reupload
`duration`	uint32	seconds, all parts combined
`pubdate`	timestamp	publish time
`ctime`	timestamp	upload time
`pic`	string	cover image url
`state`	int32	bilibili's review state, 0 is normal
`attr`	int32	raw attribute bitfield
`videos`	uint16	number of parts
`cids`	list[uint64]	part ids, in order
`page_titles`	list[string]	part titles, same order as `cids`
`fetched_at`	timestamp

comments

column	type	notes
`rpid`	uint64	comment id
`oid`	uint64	the commented object, the `aid` for videos
`type`	uint8	1 for videos
`root`	uint64	top-level comment of the thread, 0 if top
`parent`	uint64	comment being replied to, 0 if top
`mid`	uint64	commenter id
`uname`	string	commenter name at fetch time
`message`	string
`ctime`	timestamp
`likes`	uint32	counter at fetch time
`rcount`	uint32	reply count at fetch time
`state`	uint8	raw state flag
`fetched_at`	timestamp

danmaku

column	type	notes
`cid`	uint64	video part id
`dmid`	uint64	danmaku id
`progress`	uint32	position in the video, milliseconds
`mode`	uint8	1-3 scrolling, 4 bottom, 5 top, 7 advanced, 8 code
`fontsize`	uint8
`color`	uint32	rgb as an integer
`midhash`	string	sender hash, exactly as bilibili serves it
`content`	string
`ctime`	timestamp	send time
`weight`	int8	bilibili's block-level weight, 0 to 10
`pool`	uint8	0 normal, 1 subtitle, 2 special
`fetched_at`	timestamp

stats

aid, view, danmaku, reply, favorite, coin, share, likes, his_rank, fetched_at. one row per video, the latest values read.

stats_history

same counters as stats plus now_rank, with one row per reading instead of one per video. (aid, fetched_at) is unique, so a video's rows in order of fetched_at give its view, like and comment counts over time.

what this is not

not complete. videos are found through bilibili's popular lists and related videos, so popular videos and their neighbours are overrepresented. the long tail of videos nobody watches is mostly missing.
danmaku coverage depends on the video. bilibili only serves a capped pool of danmaku per video, sized by video length. for most videos that pool holds every danmaku ever sent. for videos with more than about 2000 danmaku, the full current pool is fetched instead, which still leaves out older danmaku that bilibili rotated out. very busy videos are the least complete.
engagement counts are snapshots. every counter is whatever it read at fetched_at, so do not treat these as final values. stats_history has every reading if you need the trend.
no rich user profiles. commenter ids and names are in comments, uploader ids are in videos, and that is all.

license

tiago.zip/license

disclaimers

access can be revoked at any time and all data is provided as-is. this dataset contains personal information about real people, a lot of whom are not public figures. this dataset is not a reflection of the views of the author or the institutions they are affiliated with. please be respectful. everything in this dataset was publicly readable at the time it was fetched. videos, comments and most danmaku were fetched without logging in. the full danmaku pool on busy videos was fetched with a logged-in account, which reads the same danmaku the bilibili app shows to any viewer.

Bilibili is a trademark of its respective owner. this dataset is independent and unaffiliated.

opting out

you may opt out at optout.tiago.zip.

citation

@misc{bilibili,
  title  = {bilibili: a corpus of public Bilibili videos, comments and danmaku},
  author = {tiagozip},
  year   = {2026},
  url    = {https://huggingface.co/datasets/tiagozip/bilibili}
}

Downloads last month: 725

Total file size:

293 GB