VynFi/vynfi-sar-narratives

---
license: apache-2.0
task_categories:
- text-generation
- tabular-classification
tags:
- synthetic
- financial-data
- vynfi
- aml
- sar
- banking
size_categories:
- 100K<n<1M
---

# VynFi SAR Narratives: AML Labels with Transaction Evidence

156,787 banking transactions paired with 156,714 AML labels and case-level SAR narrative text. Financial services sector, 6 monthly periods.

## Configs

| Config | Records | Columns |
|--------|---------|---------|
| `transactions` | 156,787 | 59 |
| `labels` | 156,714 | 17 |
| `narratives` | 1 | 13 |

The narrative count is low because narratives are generated per suspicious case, and the default suspicious rate is ~0.05%. For more narratives, generate with a higher `fraudRate` or `suspiciousRate` in the banking config.

## Limitations

- Narrative text is template-generated, not LLM-written. It follows a fixed structure (subject, activity summary, indicators, recommendation).
- The evidence chain join (narrative → labels → transactions) requires matching on `case_id`. Not all transactions have a case assignment.
- This dataset is a starting point for SAR NLP research, not a finished benchmark.

## Citation

```bibtex
@dataset{ivertowski_vynfi_sar_2026,
title = {VynFi SAR Narratives: AML Labels with Transaction Evidence},
author = {Michael Ivertowski},
year = {2026},
url = {https://huggingface.co/datasets/VynFi/vynfi-sar-narratives},
note = {Generated with VynFi (https://vynfi.com)}
}
```

License: Apache 2.0. Entirely synthetic.