# Dataset card - Crypto Accounting Bench

## Summary

118 crypto-accounting tasks. Each task supplies one transaction, the evidence that
surrounded it, and the complete chart of accounts of the organization that recorded it,
and asks for the full journal entry that organization actually posted.

- **Task type:** structured generation / financial reasoning
- **Languages:** English
- **Size:** 118 tasks, 7 pseudonymized organizations
- **Split:** single `test` split. This is an evaluation set, so there is no train split
- **Answer supervision:** complete journal entry, plus a frozen weighted rubric
- **License:** [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)

This is the same 118-task population reported in the research paper *Crypto Accounting
Bench: Evaluating Frontier and Open-Weight Models on Crypto-Asset Accounting Tasks*
(Entendre Finance, September 2026), transformed for public release.

## Links

- Scoring harness: `examples/run_benchmark.py`, shipped with this dataset and
  dependency-free.
- Method, metric definitions and reproduction notes: `README.md` in this dataset.
- Research paper: *Crypto Accounting Bench: Evaluating Frontier and Open-Weight Models
  on Crypto-Asset Accounting Tasks* (Entendre Finance, September 2026). It is not
  publicly distributed and not on a preprint server, so no arXiv identifier or DOI is
  asserted.

## Provenance and transformation

The tasks derive from real production accounting records. The evaluated corpus uses
pseudonymized identifying names and keeps private database identifiers outside model
inputs. This public dataset is a **transformed derivative** of the same 118 underlying
cases rather than the exact inputs used in the reported evaluation: a deterministic,
one-way transformation replaced identity-bearing and correlatable values with
synthetic ones, consistently across the whole dataset, while preserving the accounting
relationships the task depends on. Generic accounting terminology and some public
categorical labels, among them the source system's exchange enum, are published as
recorded. Debits equal credits, amounts remain derivable from
visible evidence, tax-lot and gain/loss arithmetic stay exact, related records stay
related, and name-based inference links survive.

The mapping from real values to synthetic ones is **not** contained in this directory
and cannot be derived from it. These measures do not establish complete anonymity.

Full detail: `TRANSFORMATION_REPORT.md`. Audit results: `PUBLICATION_AUDIT.md`.

## Intended use

- Evaluating whether models can infer an organization-specific accounting treatment
  from transaction evidence.
- Studying structured financial output, long-context account selection (a median of 569
  candidate accounts per task), and multi-signal reasoning.

## Out of scope

- Any use that treats a record as a factual statement about a real organization,
  person, counterparty or transaction. The records are synthetic derivatives.
- Attempting to re-identify the source organizations or transactions.
- Accounting, tax or investment advice. The expected answers reflect specific
  organizations' conventions, not general guidance, and are not authoritative.
- Training data for models intended to produce filed financial statements without human
  review.

## Composition

The set spans 8 recorded source-system transaction labels, 21 synthetic assets, 10
synthetic chains plus an exchange-only context, 2 base currencies, and the 20 legal
entities in the corpus.

| Property | Value |
|---|---|
| non-zero reference lines in total | 246 |
| flow direction | 55 inflows, 63 outflows |
| tasks with tax-lot evidence | 38 (17 created lots, 21 relieved lots) |
| tasks with a realized gain/loss line | 5 |
| tasks with related same-hash records | 60 |
| tasks with prior-receipt history for the recurring transaction pattern | 118 |
| chart of accounts shown per task | complete, a median of 569 accounts |

| Rubric family | Tasks |
|---|---|
| `TRANSFER` | 37 |
| `INCOME_EXPENSE` | 27 |
| `INTERCOMPANY` | 21 |
| `SWAP` | 16 |
| `FEE` | 12 |
| `REALIZED_GAIN_LOSS` | 5 |

Task identifiers are `task_0001` through `task_0118` with no gaps.
`manifest.json` carries the full machine-readable composition and per-file checksums.

## Known biases and limitations

- Aggregate scores characterize this evaluation population and are not a uniform sample
  of crypto accounting.
- Expected answers encode each organization's own conventions. Reasonable alternative
  treatments score as incorrect.
- Grading is keyed to exact account names within the supplied chart.
- Asset and chain identities are synthetic, so results say nothing about a model's
  knowledge of any specific real protocol.

## Citation

The paper is not on a preprint server, so no DOI or arXiv identifier is asserted. Cite
the paper and the dataset, and name the scorer that produced any result you report:

```
Kareem Khattab, Omar Khattab, and Mohamed Ibrahem.
Crypto Accounting Bench: Evaluating Frontier and Open-Weight Models on
Crypto-Asset Accounting Tasks. Entendre Finance, September 2026.
```

```bibtex
@misc{crypto_accounting_bench_2026,
  title        = {Crypto Accounting Bench: Evaluating Frontier and Open-Weight
                  Models on Crypto-Asset Accounting Tasks},
  author       = {Khattab, Kareem and Khattab, Omar and Ibrahem, Mohamed},
  year         = {2026},
  month        = sep,
  institution  = {Entendre Finance},
  note         = {Public benchmark dataset and evaluation set},
  howpublished = {\url{https://huggingface.co/datasets/Entendre/Crypto-Accounting-Bench}}
}
```

Reproductions should record the Hugging Face dataset revision they ran against, and the
scorer that produced the result.