VynFi/vynfi-sar-narratives

VynFi SAR Narratives — AML Labels with Transaction Evidence

156 787 banking transactions paired with 156 714 AML labels and case-level SAR (Suspicious Activity Report) narrative text. Designed as a starting point for SAR NLP research and for end-to-end pipelines that go from raw banking activity → AML labels → human-readable narrative.

Generated with DataSynth (banking + narrative modules) · GitHub · Companion paper (SSRN).

Provenance note. This dataset was last refreshed under DataSynth v5.0. The v5.6 → v5.9 release line ships fixes centred on the journal-entry / accounting-network generators; the banking + narrative modules producing this dataset were untouched, so the data here is unaffected by those changes and a regeneration is not in scope for the v5.9.0 fleet refresh.

What's included

Config	Rows	Columns	What it is
transactions (default)	156,787	59	One row per banking transaction (same schema as VynFi/vynfi-aml-100k).
labels	156,714	17	Per-transaction AML labels with case_id, typology, is_suspicious, severity.
narratives	1	13	Per-suspicious-case SAR narrative text (subject, activity summary, indicators, recommendation).

Quick start

from datasets import load_dataset

txns       = load_dataset("VynFi/vynfi-sar-narratives", "transactions", split="train")
labels     = load_dataset("VynFi/vynfi-sar-narratives", "labels",       split="train")
narratives = load_dataset("VynFi/vynfi-sar-narratives", "narratives",   split="train")

# Join chain: narrative → labels → transactions
import pandas as pd
n = narratives.to_pandas()
l = labels.to_pandas()
t = txns.to_pandas()
joined = n.merge(l, on="case_id").merge(t, on="transaction_id")

Generation

DataSynth release	v5.0 (banking + narrative modules — unchanged in v5.6–v5.9)
Industry	Financial services
Period	6 monthly periods
Suspicious rate	≈ 0.05 % (production-realistic)
Reproducibility	Determined by the embedded ChaCha8 seed; regenerable from a future v5.9-aware banking config (not yet pinned in configs/examples/hf/).

Limitations

• Narrative text is template-generated, not LLM-written. It follows a fixed structure (subject, activity summary, indicators, recommendation) — useful for evaluating narrative-style models against a known schema, but not a substitute for production-quality case write-ups.
• The evidence chain join (narrative → labels → transactions) requires matching on case_id. Not every transaction has a case assignment.
• The narrative count is low because narratives are produced per suspicious case, and the default suspicious rate is ~0.05 %. For more narratives, regenerate with a higher fraud_rate or suspicious_rate in the banking config.
• This dataset is a starting point for SAR NLP research, not a finished benchmark.

License

Apache-2.0.

Citation

@misc{ivertowski2026datasynth,
  author       = {Ivertowski, Michael},
  title        = {{DataSynth}: Reference Knowledge Graphs for Enterprise
                  Audit Analytics through Synthetic Data Generation
                  with Provable Statistical Properties},
  year         = {2026},
  month        = {April},
  howpublished = {SSRN Working Paper},
  url          = {https://ssrn.com/abstract=6538639}
}