Nachammai41/underserved-persona_conditioned-fraud-v2

Persona-Conditioned Fraud Detection Dataset (v2)

Overview

Synthetic fraud detection dataset for 4 underserved financial archetypes, generated using persona-conditioned sampling. Each transaction is anchored to a named persona with structured world dimensions (corridor, service loyalty, cadence, fraud-vector history, language mix), enabling behavioral coherence verification against the persona that produced it.

Quick Start

from datasets import load_dataset

# The full 20k-row dataset
ds = load_dataset("<user>/<repo>", name="all")["train"]

# One archetype only
remit = load_dataset("<user>/<repo>", name="remittance")["train"]

# Look up the persona behind persona_id="rem_004"
personas = load_dataset("<user>/<repo>", name="personas")["train"]
profile  = personas.filter(lambda r: r["persona_id"] == "rem_004")[0]

Dataset Statistics

Metric	Value
Total transactions	20,000 (5,000 per archetype)
Total personas	40 (11 remittance, 11 gig worker, 9 unbanked, 9 ITIN)
Fraud rate	~10% per archetype
Conditioning schemas	40 (expanded from personas via Adaption Labs)
Verification rounds	4 iterative rounds with coherence scoring
Best coherence pass rate	90% (ITIN archetype, Round 4)
Languages represented	10 (en, es, vi, ht, hi, yo, fr, zh, tl, pt)

Archetypes

Archetype	Personas	Key World Dimensions	Best Coherence
Remittance	11	corridor_country, transfer_service_loyalty, family_crisis_history, sender_tenure	0.589 mean (R3)
Gig Worker	11	platform_mix, daily_cashout_pattern, device_stability, sim_history	0.402 mean (R3)
Unbanked	9	kiosk_location, prepaid_card_stack, income_source, documentation_status	0.609 mean (R4)
ITIN	9	business_type, tax_filing_history, credit_file_age, accountant_relationship	0.798 mean (R4)

Available Configs

Transaction data:

• all — 20,000 rows across 4 archetypes (main dataset)
• remittance / gig_worker / unbanked / itin — 5,000 rows each

Supplementary:

• personas — 40 persona profiles (flat, one row per persona)
• conditioning_schemas — 40 expanded-world schemas (input to the generator)

Coherence verification progression:

• coherence_round1 — v1 baseline (random persona assignment), 800 rows
• coherence_round2 — v2 R1 (persona-anchored), 200 rows
• coherence_round3 — v2 R2 (cadence/fee/amount tightened), 200 rows
• coherence_round4 — v2 R3 (joint platform+hour sampling), 200 rows
• coherence_latest — alias pointing to the final round
• coherence_progression — 16-row summary (4 rounds × 4 archetypes)

Schema (per transaction row)

Field	Type	Description
data_uuid	string	Unique identifier
persona_id	string	Source persona (e.g., rem_004, gig_001) — join to personas config
archetype	string	remittance, gig_worker, unbanked, itin
dataset_version	string	"v2"
transaction_amount_usd	float	Amount in USD
fee_amount_usd	float	Fee in USD
sender_age	int	Persona-derived age with jitter
hour_of_day	int	Transaction hour (persona-window constrained)
day_of_week	int	0=Monday through 6=Sunday
day_of_week_name	string	Human-readable day
days_since_last_txn	int	Cadence-derived interval
account_age_days	int	Tenure-derived account age
txn_count_30d	int	Cadence-derived monthly count
instrument	string	Payment method (persona-specific)
language	string	Language code from persona's language_mix
fraud_vector	string	Fraud type or instrument label
is_fraud	int	0 = legitimate, 1 = fraudulent
device_type	string	Persona's device
device_stability	float	Device churn score
narrative_text	string	Adaption-generated narrative description
detected_language_hints	string	Languages detected in the narrative
fraud_vector_hint	string	Fraud pattern hint
record_timestamp	string	ISO timestamp
source	string	Provenance tag
id	string	Legacy id field

Generation Pipeline

persona_profiles.json
    └── Adaption Labs "Expand the World"
        └── conditioning_schemas/train.parquet  (40 personas × 5 world schemas)
            └── TabDDPM v2 generator (persona-conditioned sampling)
                └── transactions (5k per archetype)
                    └── Adaption Labs coherence scoring
                        └── coherence_round{1..4}/train.parquet

Coherence Verification Progression

Round	Method	Remittance	Gig Worker	Unbanked	ITIN
R1	v1 baseline (random persona)	0.090	0.168	0.145	0.097
R2	v2 persona-anchored	0.426	0.370	0.543	0.720
R3	cadence/fee/amount tightened	0.589	0.402	0.540	0.718
R4	joint platform+hour sampling	0.516	0.399	0.609	0.798

At R4, Remittance dipped in pass rate — the deterministic tightening approach hits a ceiling on event-driven archetypes where fraud vectors (phone scams, courier theft, fake-ICE calls) are discrete events rather than schedule features. Resolving this requires an agentic or small-language-model generation layer and is the target of v2_r4 / future v3 work.

Origin

This dataset was created as part of the Uncharted Data Challenge by Adaption Labs (April 2026). It extends the Fraud Detection Framework — an Agentic RAG pipeline with a custom Financial SLM built on the IEEE-CIS dataset (AUC-ROC 0.9486). The underserved dataset enables direct benchmarking: how does a model trained on mainstream data perform on populations it has never seen?

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; any resemblance to real individuals is coincidental.

Credits

• Adaption Labs — Expand World and coherence scoring
• Tab-DDPM — Gaussian diffusion for tabular data