Nachammai41/underserved-persona_conditioned-fraud-v4-cot

Persona-Conditioned Fraud Detection — CoT Reasoning Companion (v4)

A 3,926-row chain-of-thought dataset for SFT and LLM-as-judge work. Each row pairs a v4 fraud-narrative transaction with a step-by-step reasoning trace explaining how an analyst would evaluate it.

This is the companion repo to `Nachammai41/underserved-persona_conditioned-fraud-v4` (20,300-row narrative dataset + persona/source/typology references). The two are split by size: keep the main repo lean, the CoT traces here.

Headline

Quick start

from datasets import load_dataset

cot = load_dataset(
    "Nachammai41/underserved-persona_conditioned-fraud-v4-cot",
    split="train",
)

print(cot[0]["cot_reasoning_trace"])    # step-by-step reasoning
print(cot[0]["cot_completion"])         # final verdict + supporting analysis
print(cot[0]["narrative_text"])         # the v4 narrative under review
print(cot[0]["is_fraud"])               # ground-truth label

To enrich a row with full v4 context (persona profile, source citations, typology definitions), load the parent dataset and join on data_uuid:

parent = load_dataset(
    "Nachammai41/underserved-persona_conditioned-fraud-v4",
    name="all", split="train",
).to_pandas().set_index("data_uuid")

cot_df = cot.to_pandas()
enriched = cot_df.merge(parent, left_on="data_uuid", right_index=True,
                        how="left", suffixes=("", "_parent"))

Schema

Reasoning traces average ~6 KB per row.

How the pairs were selected

For each of the 1,963 fraud rows in the parent v4 dataset:

1. 1.Find non-fraud candidates in the same archetype and same instrument, within the same log-spaced amount_band.
2. 2.If no strict triple match, relax to (archetype, amount_band).
3. 3.Pick one without replacement; record the pairing.

887 of 1,963 fraud rows fell back to the relaxed (archetype, amount_band) match — still 2 of 3 dimensions matched. The result is a roughly balanced "hard negatives" set: each fraud row sits next to a non-fraud row that looks similar on the structured features.

Suggested uses

• —SFT for fraud-analyst LLMs — train a model to emit reasoning traces given (transaction metadata + narrative) → verdict.
• —LLM-as-judge fine-tuning — distill the reasoning style into a smaller model used in evaluation pipelines.
• —CoT data augmentation — combine with the parent's 20,300-row bundle for mixed reasoning + narrative-only training.
• —Hard-negatives evaluation set — the 1:1 matched pairs make a clean test bed for measuring how well a fraud detector handles look-alike legitimate transactions.

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

CC-BY-4.0. Same provenance and disclaimers as the parent dataset: synthetic transactions about fictional personas, narratives and reasoning traces are LLM-generated, no real persons or real fraud cases involved.

Credits

• —Adaption Labs — reasoning_traces recipe (the source of these traces)
• —Parent dataset (citation-grounded personas + typology + sources): `Nachammai41/underserved-persona_conditioned-fraud-v4`
• —Underlying source registry, FinCEN typology, and persona grounding documented in the parent repo's README.