Botijada: fake and coordinated followers on Instagram (Serbia)

The goal of this dataset is to help detect fake and coordinated followers.

It contains follower lists from Instagram profiles in Serbia that were suddenly followed by thousands of unknown accounts. The profile owners downloaded the lists from Instagram themselves and sent them to botijada.surf. Next to the raw follows are the results of the tests the site uses to find accounts that act together.

All usernames, profiles and lists are replaced with random IDs. The same account keeps the same ID in every release. There are no names, contacts or free text.

Srpski: Cilj ovog skupa podataka je da pomogne u otkrivanju lažnih i koordinisanih pratilaca. Sadrži spiskove pratilaca koje su vlasnici napadnutih Instagram profila u Srbiji sami preuzeli sa Instagrama i poslali sajtu botijada.surf. Sva korisnička imena, profili i spiskovi zamenjeni su nasumičnim oznakama.

Tables

Each table is a separate config with one train split. There is no test split. If you make one, split by target_id or by time, not by row, so the same account doesn't end up on both sides.

Table	One row is
follow_actions	an account following a target (or asking to, for private targets). The analyses run on this.
follows	one continuous follow, with the lists that saw it start and stop
sightings	an account as shown in one list
requests	a follow request shown in one list (private targets)
targets	a profile that sent lists: its kind and how publicly active it is
lists	a list that was sent: when, which period it covers, which signs the owner reported
hourly	new follows of a target in one UTC hour
spikes	a sudden jump in new followers
target_analysis	a target's usual rate of new followers and its current state
networks, network_members, network_targets	groups of accounts that followed the same profiles within minutes, their members, and the profiles they hit
steady_accounts, steady_chains	accounts that move from one profile to the next within minutes, and their paths
pairs	a test of two targets: did their shared followers follow both close together more often than chance
campaign_networks, campaign_pairs	whether two networks hit the same profiles more often than chance
codebook	labels for the coded values, in Serbian and English

manifest.json has the row counts, all analysis parameters, and the numbers shown on the site for the same data. SHA256SUMS has a checksum for each file.

IDs and times

• account_id (a_ + 16 hex characters), target_id (T_ + 8) and list_id (L_ + 10) are random. They are not derived from usernames, so nobody can compute one from a username. The mapping stays private on the project's server.
• network_id (N1, N1.W2) is numbered by size and only holds within one release.
• Times are in UTC, to the second.
• We don't know when an account was removed. We only know the first later list that no longer shows it (gone_list_id).
• Target kinds with fewer than 3 profiles are grouped as other.

How the data was collected

Profile owners used Instagram's "Download your information" and sent the follower files (and, for private accounts, the follow requests file) through the form on botijada.surf, agreeing to this use. Nothing was scraped. Many people sent several lists over a few days, which is how we can tell when accounts disappear.

How coordination is tested

Each test compares the data with a shuffled version that keeps every profile's follow times exactly as they were (its spikes, viral days and daily rhythm) and only mixes up which account holds which time. Two profiles that went viral on the same day look normal under this test. What stands out is the same accounts following several profiles close together.

• Pairs: accounts that followed both targets within 4 hours, compared with the shuffle, at a 1% false discovery rate.
• Networks: two accounts are linked when they followed the same profiles within 5 minutes more often than chance. Linked accounts form networks, which are split into waves.
• Steady accounts: a single account moving from profile to profile within minutes, again and again.
• Shared targets: two networks hitting the same profiles more often than their size explains.

The networks and steady accounts can be recomputed exactly from follow_actions (rows in ord order, targets in target_ord order). Pair counts also match exactly. Permutation p-values match up to random error.

Limits

• Only profiles whose owners sent lists are included. This is not a sample of Instagram.
• The share of removed accounts is a lower bound. Accounts removed before the first list or after the last one are never seen.
• A list that no longer shows an account doesn't say why. Instagram may have removed it, it may have unfollowed, the owner may have removed it, or it changed its name.
• The earliest lists (has_files is false) were sent before the original files were kept, so removals in that period are not dated precisely.
• A renamed account gets a new ID.
• Acting together is not proof that an account is fake, or of who runs it.

Privacy and use

The IDs are pseudonymous, not anonymous. Follow times are exact and public follower lists can be seen on Instagram, so someone who knows a target could match some IDs back to usernames.

Please don't:

• try to identify accounts or targets;
• harass or report anyone based on this data;
• use it to build tools for buying followers or avoiding detection.

The license (CC BY 4.0) doesn't forbid these uses, so this is a request.

To have an account removed, open a discussion in this repository's Community tab and ask us to contact you. Don't post the username there. We remove the account, publish a new release within 30 days and rewrite this repository's history. Copies that were already downloaded can't be recalled.

Versions

Releases are tagged by month (v2026.10, then v2026.11 and so on). IDs stay the same between releases, except network_id. Changes are listed in CHANGELOG.md.

Quickstart

from datasets import load_dataset
follows = load_dataset("gavrilo/botijada", "follow_actions", revision="v2026.10")["train"].to_pandas()

-- DuckDB
SELECT target_id, count(*) FROM 'hf://datasets/gavrilo/botijada/data/follow_actions/*.parquet' GROUP BY 1;

Citation

@misc{botijada_2026,
  title  = {Botijada: fake and coordinated followers on Instagram (Serbia)},
  author = {{botijada.surf}},
  year   = {2026},
  note   = {Release v2026.10},
  url    = {https://huggingface.co/datasets/gavrilo/botijada}
}

Made with SHARE Fondacija. Source: botijada.surf, follower lists sent by profile owners.