Datasetpublic

Etherlabs/ios-risk-finetune-v3

IOS Risk Fine-Tune Dataset v3 Quality-gated instruction-tuning data for financial fraud, AML typologies, and Bank Secrecy Act regulatory recall. This is a research dataset assembled from public data, official public regulations, deterministic synthetic scenarios, and validated model-assisted rewrites. It is not production transaction evidence. Composition Source Records Description Public tabular benchmark 9,242 ULB/Kaggle credit-card examples; record… See the full description on the dataset page: https://huggingface.co/datasets/Etherlabs/ios-risk-finetune-v3.

sourceHugging Faceapache-2.0updated 29d agoView on Hugging Face

0likes64downloads

Dataset Card

IOS Risk Fine-Tune Dataset v3

Quality-gated instruction-tuning data for financial fraud, AML typologies, and Bank Secrecy Act regulatory recall. This is a research dataset assembled from public data, official public regulations, deterministic synthetic scenarios, and validated model-assisted rewrites. It is not production transaction evidence.

Composition

Source	Records	Description
Public tabular benchmark	9,242	ULB/Kaggle credit-card examples; record holdout enforced
AML typology cases	6,000	Six typologies plus documented benign counterexamples
Fraud scenarios	5,000	Card testing, takeover, mule, bust-out, and benign cases
eCFR BSA regulations	364	Official 31 CFR Chapter X text; whole sections held out

The AML set includes 2,462 strictly validated rewrites produced with nvidia/nemotron-3-super-120b-a12b; the remaining AML answers use deterministic templates. Generated records were validated before inclusion.

Format

json

{
  "instruction": "Classify this transaction as FRAUD or LEGITIMATE.",
  "input": "Amount: $149.62 | Hour: 0 | OffHours: 1 | ...",
  "output": "LEGITIMATE",
  "source": "ulb_tabular_train"
}

Usage

python

from datasets import load_dataset

dataset = load_dataset("Etherlabs/ios-risk-finetune-v3")

Quality and leakage controls

—20,606 unique instruction/input pairs
—source-record hash partition for tabular train/test separation
—whole-section partition for regulatory train/test separation
—malformed, duplicate, fallback, and unsupported-number rewrite rejection
—independently authored counterfactual risk evaluation cases kept outside training

Limitations

—Most risk scenarios are synthetic and cover a bounded set of patterns.
—Regulatory pairs teach recall of text and citations, not legal judgement.
—The tabular benchmark lacks rich device, merchant, customer-history, and graph features.
—Do not use a downstream model as an autonomous fraud, account-restriction, or regulatory-filing decision-maker.
—Compare any fine-tune with its untouched base model on the held-out Project 03 evaluation before making an improvement claim.