SavienWalker/bulwark-bench-tr

Bulwark Bench TR

A Turkish benchmark for scam / spam / legitimate message classification, built to measure what existing detectors get wrong: real Turkish institutional messages (bank, cargo, e-Devlet, operator, one-time codes) and hard scams that imitate them.

Model: SavienWalker/bulwark-scam-detector · Code (collection, masking, labelling, training): github.com/SavienWalker/bulwark

🇹🇷 Türkçe özet: Gerçek bir telefonun gelen kutusundan 434 SMS, 40 e-posta ve sahte–gerçek karşıt çiftler içeren 102 zor mesaj. Bütün kişisel veriler maskelenmiştir. Amaç, dolandırıcılık dedektörlerinin gerçek kurum mesajlarına verdiği yanlış alarmı ve taklit mesajları ayırma becerisini ölçmektir.

Subsets

Config	Messages	Source	Labels (legit / spam / fraud)	Notes
real_sms	434	one person's iPhone inbox, 2024–2026	193 / 112 / 129	eval_split: 222 dev, 212 test; 125 ambiguous
real_email	40	same person's Gmail inbox	21 / 17 / 2	used as development data for Bulwark
real_other	1	screenshot of a fake reward-survey page	0 / 0 / 1	non-message channel example
hard_synthetic_v1	42	written by hand, LLM-assisted	17 / 4 / 21	7 contrast pairs
hard_synthetic_v2	60	written by an LLM (Claude), human-reviewed	20 / 0 / 40	15 contrast pairs, 5 English

Recommended use: report real_sms (test split) as the main real-world number, and the synthetic sets separately. real_sms dev split and real_email were used to develop Bulwark; if you compare against Bulwark, use the test split.

Fields

Field	Description
text	message text, personal data masked
label	legitimate, spam (ad from a real business) or fraud (tries to steal money, data, codes or trust)
category	topic: banka, kargo, devlet, otp_kod, bahis, kampanya, is_ilani, yatirim, …
difficulty	kolay / orta / zor: how likely an average person is to be fooled (fraud) or a model to raise a false alarm (legitimate)
hard_tags	observable traits: link_yok, benzer_alan_adi, gercek_alan_adi, kisaltilmis_link, resmi_dil, sohbet_tonu, gercek_kurum_bildirimi, aciliyet, para_tutari, kod_iceriyor, …
explanation	short reason why the message is fraud/spam (Turkish)
pair	shared key linking a fake message to the genuine message it imitates
ambiguous	true → exclude from headline metrics (mostly illegal-betting ads)
channel	sms, imessage, eposta, whatsapp, telegram, sosyal_medya, diger
sender, sender_type	alphanumeric sender name (e.g. a bank) or just the type (numara, kisa_kod); phone numbers are removed
labeled_by	insan (labelled by a person) or claude_onayli (suggested by Claude, approved one by one by a person)
eval_split	dev / test, split by sender so one institution never appears in both
source, is_synthetic, date	provenance; dates are month-level only

How it was built

• Real SMS: exported from an encrypted iPhone backup. Outgoing messages, group chats and anyone the owner had replied to (friends, family) were excluded. Near-duplicate templates were collapsed (max. 3 messages per institution) so a single bank cannot dominate the scores.
• Masking: phone numbers, IBANs, card numbers (Luhn-checked), national ID numbers (checksum-validated), e-mail addresses, account numbers, plates, IP addresses, URL path tokens, names after salutations or in bank-transfer notices, and the inbox owner's own names and usernames are replaced with neutral placeholders (Ad Soyad, 0500 000 00 00, TR00 …). Free text was also reviewed by hand.
• Labelling: 89 SMS labelled directly by the inbox owner; the rest suggested by Claude and approved message-by-message by the owner (labeled_by). Note that the owner accepted most suggestions unchanged.
• Hard sets: each fake message is paired with the genuine message it imitates (same topic, different domain/request/tone). A pair counts as solved only if both are classified correctly.

Baseline results (unwanted = p(spam) + p(fraud))

Set	Metric	Base model	Bulwark
real_sms test	false alarm on legitimate ↓	53% (t=0.5)	14% (t=0.6)
real_sms test	unwanted caught ↑	100%	98%
hard_synthetic_v2	fraud caught ↑	82%	95%
hard_synthetic_v2	contrast pairs fully correct ↑	9/13	12/13
hard_synthetic_v1	contrast pairs fully correct ↑	5/7	7/7

Limitations

• Single inbox. Real messages come from one person in Turkey, so institutions and ad senders reflect that person's life. Real non-betting fraud is rare (4 messages); fraud recall relies on the synthetic sets.
• Same author for hard sets and Bulwark's training templates (Claude). Style overlap may favour Bulwark on the synthetic sets; the real_sms test split is the unbiased comparison.
• Single messages only. Phone-call, multi-turn chat and website-based scams are out of scope (one website example is included in real_other).
• Personal data has been masked carefully, but if you find anything that identifies a person, please open a discussion and it will be removed.

License

CC BY 4.0. The inbox owner contributed their own messages for this purpose.