--- language: - hi - en - bn - mr - ta - te - gu - kn - ml - or - pa - ur - as - brx - doi - kok - ks - mai - mni - ne - sa - sat - sd license: other configs: - config_name: classifier data_files: - split: train path: classifier/train-*.parquet - split: validation path: classifier/validation-*.parquet - split: test path: classifier/test-*.parquet - config_name: chatml data_files: - split: train path: chatml/train-*.parquet - split: validation path: chatml/validation-*.parquet - split: test path: chatml/test-*.parquet --- # BSIS: Indian Scam Message Detection Dataset Synthetic training data for on-device scam detection on Indian SMS and WhatsApp traffic. Every row was produced by the BSIS red-team/blue-team simulation arena: a scammer (or a genuine sender) talks to a simulated victim in an Indian language, and a privacy-preserving defender scores each message. Labels come from deterministic Rust logic over ground truth, never from model opinion. This dataset is synthetic. It contains no real user messages, and no real personal data. ## Splits | config | train | validation | test | |---|---|---|---| | classifier | 32858 | 2148 | 1883 | | chatml | 28957 | 1905 | 1607 | Splits are grouped by lineage. All mutated generations of one scenario stay in the same split, so no adversarial sibling leaks across train, validation or test. ## Label balance (classifier) | label | rows | |---|---| | 0 (safe until this round) | 14598 | | 1 (scam detectable at this round) | 22291 | Label 1 marks the sender turns from which a careful human could tell the message is a scam (judge `earliest_detectable_round`). Earlier turns of the same scam carry label 0 and teach the detector not to false-alarm during grooming. ## Languages All 23 Sarvam-supported Indian languages: 22 Scheduled languages plus Indian English. Odia uses the `od-IN` code. | language | train | validation | test | |---|---|---|---| | as-IN | 1017 | 88 | 89 | | bn-IN | 2322 | 144 | 153 | | brx-IN | 893 | 83 | 61 | | doi-IN | 882 | 72 | 58 | | en-IN | 1866 | 135 | 118 | | gu-IN | 1309 | 83 | 73 | | hi-IN | 4654 | 257 | 224 | | kn-IN | 1371 | 64 | 68 | | kok-IN | 905 | 69 | 60 | | ks-IN | 857 | 62 | 51 | | mai-IN | 890 | 76 | 56 | | ml-IN | 1233 | 48 | 72 | | mni-IN | 796 | 73 | 50 | | mr-IN | 2208 | 109 | 101 | | ne-IN | 877 | 70 | 51 | | od-IN | 1139 | 78 | 25 | | pa-IN | 1315 | 87 | 67 | | sa-IN | 859 | 75 | 45 | | sat-IN | 917 | 77 | 59 | | sd-IN | 1057 | 83 | 63 | | ta-IN | 2117 | 127 | 128 | | te-IN | 2103 | 136 | 142 | | ur-IN | 1271 | 52 | 69 | ## Vectors Nine scam playbooks grounded in RBI BE(A)WARE and I4C advisories, plus four hard negatives (genuine senders). Hard negatives make up roughly a quarter of lineages, so the detector learns not to flag ordinary bank alerts, doorstep delivery OTP requests, bill reminders and family messages. | vector | train | validation | test | |---|---|---|---| | digital_arrest | 3077 | 117 | 270 | | fake_courier_customs | 2836 | 127 | 114 | | fake_customer_care | 3227 | 233 | 232 | | kyc_utility_panic | 2568 | 96 | 166 | | legit_bank_alert | 1646 | 108 | 66 | | legit_delivery_otp | 1485 | 71 | 63 | | legit_electricity_reminder | 2025 | 98 | 73 | | legit_family_transfer | 1700 | 108 | 107 | | predatory_loan_app | 2472 | 349 | 94 | | qr_code_refund_trap | 3346 | 237 | 145 | | social_friend_emergency | 2813 | 237 | 105 | | upi_collect_trap | 2771 | 136 | 197 | | whatsapp_stock_task | 2892 | 231 | 251 | ## Schema ### classifier - `text` (string): the message window, all sender turns up to and including the current one, prefixed `SENDER:` / `RECIPIENT:`. - `label` (int64): 1 if the scam is detectable at this round, else 0. - `language` (string): BCP-47 code, for example `hi-IN` or `od-IN`. - `script_mode` (string): `native_script` or `romanized_codemix`. - `vector` (string): the scam playbook or legitimate scenario name. - `round` (int64): the sender turn number. - `lineage_id` (string): UUIDv5 of the scenario lineage. Rows sharing a lineage belong to the same conversation family. - `hard_example` (bool): the defender missed, was late, or false-alarmed. ### chatml - `messages` (list of {role, content}): a three-message conversation: a short detector system prompt, the user window, and a strict-JSON assistant reply with `scam_probability`, `tactics` and `warning`. Positive windows without an agreeing on-device warning are excluded from chatml and kept in classifier only, so the assistant target never invents a warning. ## Provenance and safety - Synthetic only. Generated on the Sarvam Token Factory platform with red-team prompts restricted to placeholders: `+91-98XXX-XXXXX`, `http://example-.invalid`, `name@fakeupi`, `XXXX XXXX 1234`. - Every generated message passed a regex PII scrubber before storage, and this package re-checks each exported row for raw Indian mobile numbers, Aadhaar patterns, PAN patterns and non-`.invalid` URLs. The check passes with zero hits. - A deterministic script check rejected conversations written in the wrong script before verification: Romanized rows stay ASCII, native rows match the expected Unicode block per language. - A held-out set of real, anonymized messages is kept separately and is not part of this package. ## Use ```python from datasets import load_dataset dataset = load_dataset("harshcasper/bsis", "classifier") ``` ## License This dataset is released for research and educational use in scam detection and cyber-safety training. It is synthetic and carries no warranty. Do not use it to contact anyone or to send any message. If you need a different license, open an issue in the source repository.