Nachammai41/underserved-persona_conditioned-fraud-v4

Persona-Conditioned Fraud Detection Dataset (v4 + v4.1, Full Typology Coverage)

A 20,300-row citation-grounded synthetic fraud-narrative dataset for four underserved US financial-system archetypes — remittance, gig_worker, unbanked, ITIN — with all 25 FinCEN typology codes exercised.

What's new vs v3

V3 covered 10 of 25 FinCEN typology codes. v4 closed the gap to 18/25 through three targeted changes:

1. 1.16 persona edits documenting fraud events (SIM-swap, BEC, hawala/IVTS, hurricane scams, money-mule, synthetic-ID, unauthorized-ACH, lax-KYC, false-chargeback, COVID-imposter). 8 grade upgrades D→B.
2. 2.SAR-preference resolver picks the more-specific SAR Advisory code when both an FTA Identity code and a SAR Advisory match a fraud_vector.
3. 3.Combined 20k Adaption narrative fill — single job, 20 native scripts, per-row persona/transaction context.

v4.1 closed the remaining gap to 25/25 through:

1. 1.Strip prompt-tag leakage from 1,235 narratives (Adaption was echoing the upload metadata block into completions for ~6% of rows).
2. 2.Re-stamp 4 shadowed FTA codes (T4, T6, T9, T11) on a half-cap subset so both the FTA and SAR equivalents carry rows. No Adaption credits.
3. 3.Add 3 truly-missing codes (T7 Abuse of Access, T8 Refusal to Cooperate, SARHUMANTRAFFICKING) via persona-event additions on unb_001, gig_001, itin_010 + 300 synthetic transactions narrated by Adaption.
4. 4.Refill 51/52 empty narratives by resubmitting the original prompts.

Quick start

from datasets import load_dataset

repo = "Nachammai41/underserved-persona_conditioned-fraud-v4"  # TBD

ds       = load_dataset(repo, name="all")["train"]                  # 20,300 rows
remit    = load_dataset(repo, name="remittance")["train"]           #  5,000 rows
gig      = load_dataset(repo, name="gig_worker")["train"]           #  5,100 rows
unbanked = load_dataset(repo, name="unbanked")["train"]             #  5,100 rows
itin     = load_dataset(repo, name="itin")["train"]                 #  5,100 rows
personas = load_dataset(repo, name="personas")["train"]             #     46 rows
sources  = load_dataset(repo, name="sources")["train"]              #     13 rows
typology = load_dataset(repo, name="typology_registry")["train"]    #     25 rows

Dataset Statistics

Narrative quality

Combined 20k Adaption fresh fill recovers row-level signals that an overlay approach would have lost:

Archetypes

Available Configs

Transaction data:

• —all — 20,300 rows across all 4 archetypes
• —remittance (5,000) / gig_worker (5,100) / unbanked (5,100) / itin (5,100)

Reference / attribution:

• —personas — 46 persona profiles with grade, source IDs, familycrisishistory
• —sources — 13-entry citation registry
• —typology_registry — 25 FinCEN typology codes with appliestofraud_vectors

SFT companion (separate repo): see `Nachammai41/underserved-persona_conditioned-fraud-v4-cot` — 3,926 rows of chain-of-thought reasoning traces.

Schema

Same 25 columns as v3, plus three universal grounding columns (persona_source_ids, fraud_vector_typology_ref, behavioral_evidence_grade). See dataset_card.md for field-by-field descriptions.

V4 + v4.1 Typology Coverage Table (all 25 codes)

Total fraud rows: 2,263. All 25 codes carry rows.

Companion: CoT reasoning dataset (separate repo)

A 3,926-row chain-of-thought dataset (1,963 fraud + 1,963 matched non-fraud) is published as a separate repo to keep this one lean:

`Nachammai41/underserved-persona_conditioned-fraud-v4-cot`

Each row pairs a v4 transaction with a step-by-step reasoning trace produced by Adaption's reasoning_traces recipe (graded E → A, +92%). Use it for SFT, LLM-as-judge distillation, or CoT data augmentation. Rows join back to this dataset's all config on data_uuid.

Citation

@dataset{palaniappan2026underserved,
  author    = {Palaniappan, Nachammai},
  title     = {Underserved Financial Fraud Dataset},
  year      = {2026},
  publisher = {HuggingFace},
  note      = {Created with Adaptive Data by Adaption.
               Uncharted Data Challenge, Adaption Labs.},
  url       = {https://huggingface.co/datasets/nachammai779/underserved-financial-fraud}
}

License

Released under CC-BY-4.0 for research and educational purposes. Persona names are fictional; biographical details are composed from published aggregate source evidence. Any resemblance to real individuals is coincidental.

Credits

• —Adaption Labs — narrative fill (combined 20k v4 job + v4.1 patch + CoT job)
• —FinCEN — Financial Trend Analysis 2024 (Identity), SAR Advisory Key Terms
• —FDIC — 2023 National Survey of Unbanked and Underbanked Households
• —Menjívar, Agadjanian & Oh — "The Contradictions of Liminal Legality" (Soc Probl 2022)
• —Del Real — "Seemingly inclusive liminal legality" (J Ethn Migr Stud 2022)
• —Vallas & Schor — "What Do Platforms Do?" (Annu Rev Sociol 2020)
• —Remitly, Wise, Inter-American Dialogue, Oxfam America, IRS SOI, Treasury OIG, Federal Reserve FedPayments Improvement — industry & regulatory sources
• —Tab-DDPM — Gaussian multinomial diffusion for tabular data