This dataset was collected from public Bilibili web APIs by the Zeroroku observatory (zeroroku.com). It is released for non-commercial research only. By requesting access you agree not to use it to identify, profile or contact individual users, and to comply with Bilibili's terms of service and the laws that apply to you.

Bilibili Creator Stats

Time series of follower counts and video engagement for Bilibili (哔哩哔哩) creators, collected by the Zeroroku observatory (06 数据观测站), mostly between 2019 and 2026, plus the video, creator and tag metadata needed to interpret them.

Subset	Rows	What one row is
`video_stats`	2.66 B	one crawl of one video's counters
`author_fans`	1.70 B	one crawl of one creator's follower count
`video_info`	111 M	one video (latest known metadata)
`author_info`	12.0 M	one creator (latest known profile)
`tags`	12.0 M	one video tag
`live_paid_events`	~63 M	one paid live event: gift, guard purchase or super chat
`live_room_gifts`	~63 M	one paid gift in a live room
`live_paid_event_aggregations`	~32 M	paid amount of one live room in one 5-minute bucket
`livers`	~0.3 M	one live streamer and room (latest known state)

video_stats covers 108.8 M distinct videos and author_fans 11.1 M distinct creators. Counts are for the first release, v2026.09.30.

Usage

from datasets import load_dataset

fans = load_dataset("Jannchie/bilibili-creator-stats", "author_fans", split="train", streaming=True)

The files are plain Parquet, so DuckDB or Polars can query them in place. Every file in the time-series subsets is sorted by (id, crawled_at), so filtering on one aid / mid only reads a few row groups:

SELECT crawled_at, fans
FROM 'hf://datasets/Jannchie/bilibili-creator-stats/data/author_fans/**/*.parquet'
WHERE mid = 546195
ORDER BY crawled_at;

Subsets

All timestamps are UTC. Join keys: aid (video), mid (creator), tag_id.

`video_stats`

Column	Type	Notes
`aid`	int64	video id
`crawled_at`	timestamp	when the counters were read
`view`, `danmaku`, `reply`, `favorite`, `coin`, `share`, `like`	int64	cumulative counters at `crawled_at`

`author_fans`

Column	Type	Notes
`mid`	int64	creator id
`crawled_at`	timestamp	when the count was read
`fans`	int64	follower count; `null` marks a placeholder row for a newly tracked creator

`video_info`

aid, bvid, mid (uploader), title, desc, tid / tname (category), copyright, pubdate, ctime, duration (seconds), videos (page count), state, attribute, cid, pic (cover URL), dynamic, mission_id, tag_list (tag ids, see tags), dimension / rights / stat (raw JSON strings from the API), first_seen_at, updated_at.

`author_info`

mid, name, face (avatar URL), sign, sex, level, rank, silence, join_time, official / vip / live_room (raw JSON strings), is_deleted, is_senior_member, top_photo, first_seen_at, updated_at.

`tags`

tag_id, tag_name, content, short_content, type, state, ctime, view, use, atten, likes, hates, cover, head_cover, updated_at.

Live streaming subsets

About 60 k live rooms were monitored over Bilibili's live websocket. Amounts are in CNY.

live_paid_events (from 2026-01): event_type (SEND_GIFT, GUARD_BUY, USER_TOAST_MSG[_V2], SUPER_CHAT_MESSAGE[_JPN]), room_id, sender_hash, receiver_uid / receiver_name (the streamer), amount (unit price for every event type; the paid total is amount * count, where count is the number of gifts or guard months), count, gift_id, gift_name, guard_level, guard_role_name, guard_count, op_type, unit, message (super chat text), start_time, end_time, created_at. One purchase is one row: Bilibili announces a super chat twice (SUPER_CHAT_MESSAGE and SUPER_CHAT_MESSAGE_JPN) and a guard purchase up to three times (GUARD_BUY, USER_TOAST_MSG[_V2]), and the repeats are dropped. A SUPER_CHAT_MESSAGE_JPN row is a chat whose SUPER_CHAT_MESSAGE notice was missed.
live_room_gifts (from 2026-01): room_id, gift_id, price (unit price), count, sender_hash, created_at. Free gifts and guard purchases are not recorded; guards are in live_paid_events.
live_paid_event_aggregations (from 2022-08): bucket_start, room_id, gift_amount, guard_amount, sc_amount, total_amount, updated_at. From 2026-01-13 on, the three amounts are sums over live_paid_events. Buckets before 2026-01-13 come from an older collector that summed gifts, guard purchases and super chats into one figure, kept here as gift_amount (guard_amount and sc_amount are 0); it did not keep individual events, so the figure cannot be split by type. That collector multiplied a multi-month guard purchase's total by its month count again. Where the collector stored such a purchase as a row of its own it is recognisable and was corrected (6.8 k buckets, 64.7 M CNY of excess, about 4 % of the pre-2026 total); where it was merged with other payments the excess remains, so isolated large spikes before 2026 deserve suspicion. The older collector also wrote some rows twice; those duplicates cannot be told apart from real repeats and remain in the sums.
livers: mid, room_id, name, face, head_pic, guard_num, gid, live_status, live_time, updated_at.

Viewer privacy. Senders are ordinary viewers. Their uid is replaced by sender_hash, a keyed hash (first 16 hex chars of SHA-256 over a secret key and the uid): the same viewer has the same hash across all files and releases, but the hash cannot be mapped back to an account. Sender nicknames and payment flow ids are not published. Super chat message text is published as sent; it is user-written and may itself mention people.

Updates and versions

The crawler is still running, and the dataset grows with it:

Monthly, new rows of video_stats, author_fans, live_paid_events and live_room_gifts are appended as data/<subset>/increments/<cutoff>.parquet (cutoff time in UTC, e.g. 20261101T190000Z), holding every row crawled since the previous increment up to that time. Files are not rewritten except for data corrections (see below), so a row appears exactly once across the base files and the increments.
Quarterly, video_info, author_info, tags, livers and live_paid_event_aggregations are re-exported in full, replacing the previous snapshot.

Corrections: in v2026.10.05 the live subsets were re-exported after duplicated super chat and guard notices were removed, guard amounts were normalised to unit prices, gift totals since 2026-01-13 were recomputed with count (they had summed unit prices), and legacy multi-month guard buckets were fixed. Earlier tags keep the uncorrected live data.

Every release is tagged vYYYY.MM.DD. To reproduce a result, pin the tag:

load_dataset("Jannchie/bilibili-creator-stats", "author_fans", revision="v2026.09.30")

How it was collected

A crawler polled Bilibili's public web APIs for a tracked set of creators. The polling interval depends on the creator's size and growth: creators with ~10 M followers were read every few hours, creators with ~1 k followers every few weeks. The sampling is therefore irregular and uneven — resample before comparing creators. The tracked set grew over time, mainly by following the "following" lists of already-tracked creators, so it is biased toward creators connected to popular ones and is not a random sample of Bilibili.

Known gaps and caveats

video_stats has no rows from mid-2023 to the end of 2025.
From 2026 on, video_stats only has view; the other counters are null because the endpoint used since then only returns play counts.
Both time series are dense from 2019 on; the sparse rows before that come from backfills.
Metadata subsets hold the latest known state only, not their history.
Deleted or private videos and accounts keep their last crawled values.
Until September 2026, records whose text contained a NUL character failed to be stored, so a small number of profiles and videos are missing.
Live monitoring was down on 2026-09-30 from 08:39 to 09:17 (UTC+9); live events in that window are missing.
No ordinary gifts (SEND_GIFT) since 2026-09-10 08:45 UTC: Bilibili stopped delivering them to the monitor's connections, so gift_amount is 0 from then until the gap is closed. Guards and super chats are unaffected.

Not included: live danmaku, the follow graph, and anything about Zeroroku's own site users.

License and responsible use

Released under CC BY-NC 4.0 for non-commercial research. The underlying content belongs to Bilibili and its users; this dataset is not affiliated with or endorsed by Bilibili. Creator profiles are public, but please do not use this data to identify, profile or contact individuals. For takedown requests, open a discussion on this repository.

Citation

Please cite the release tag you used in the note or version field.

@misc{jannchie2026bilibilicreatorstats,
  author       = {Jannchie},
  title        = {Bilibili Creator Stats: Follower and Video Engagement Time Series of Bilibili Creators},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Jannchie/bilibili-creator-stats}},
  note         = {Collected by the Zeroroku observatory (zeroroku.com), 2019--2026}
}

Downloads last month: 115

Total file size:

61.3 GB