--- license: cc-by-4.0 language: [en, hi] task_categories: [text-generation, token-classification] tags: [lexsi-hackathon, synthetic, upi, banking, hinglish, structured-output] pretty_name: Indian bank & UPI SMS to JSON (synthetic, trilingual) configs: - config_name: default data_files: - {split: train, path: data/train_curated.jsonl} - {split: dev, path: data/dev.jsonl} - {split: test_id, path: data/test_id.jsonl} - {split: test_ood, path: data/test_ood.jsonl} - {split: challenge, path: data/challenge_set.jsonl} --- # Indian bank & UPI SMS → JSON (English · Hindi · Hinglish) Synthetic SMS with **exact** labels (a field is labelled only if its text survived generation), plus a 44-example **hand-written challenge set**. `test_ood` uses message templates that never appear in training. Bank names are fictional; messages contain no links or phone numbers. Schema: `["is_transaction", "type", "amount", "currency", "account_last4", "instrument", "counterparty", "date", "reference", "balance", "status"]`. Non-transactions (OTP, offers, reminders, future debits, balance info, money requests) → `is_transaction: false`, all other fields null. Also included: `results/` (every model prediction, metrics, curation report from CuratorKIT, AuditKIT run records) and `figures/`.