---
license: cc-by-4.0
language:
- pl
pretty_name: Factual State Discovery Benchmark
task_categories:
- question-answering
tags:
- legal
- tax-law
- polish
- fact-elicitation
- conversational
- benchmark
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: easy
    path: data/easy.parquet
  - split: hard
    path: data/hard.parquet
---

# Factual State Discovery Benchmark

Dataset for the **Factual State Discovery Benchmark: Evaluating Fact Elicitation
in Polish Tax Law** (ACL 2026 SRW). It evaluates whether conversational agents
can systematically elicit, through dialogue, all the facts of a taxpayer's
situation from a real Polish tax interpretation document.

Each sample pairs a **factual state** (a narrative of the taxpayer's situation,
in Polish) with its decomposition into **atomic facts** — independent,
verifiable claims that serve as the ground truth a discovery agent must recover.

Benchmark code (runner, QA agent, semantic scorer): **https://github.com/pwr-ai/fsdbench**

## Splits

| Split | Samples | Atomic facts | Facts/sample (mean) | Length chars (mean) |
| ----- | ------: | -----------: | ------------------: | ------------------: |
| `easy` | 250 | 7,132 | 28.5 | 2,081 |
| `hard` | 250 | 25,742 | 103.0 | 8,216 |
| **all** | **500** | **32,874** | | |

The split boundary is the factual-state length (`easy` ≤ ~4,500 chars, `hard`
above). Samples span IP Box, VAT, cryptocurrency, mortgage settlements, and
employment-tax scenarios.

## Schema

| Field | Type | Description |
| ----- | ---- | ----------- |
| `id` | int64 | Source tax-interpretation document id |
| `sample_idx` | int64 | Original sample index in the source corpus |
| `factual_state` | string | Narrative of the taxpayer's situation (Polish) |
| `atomic_facts` | list&lt;string&gt; | Ground-truth atomic facts derived from the narrative |
| `factual_state_length` | int64 | Character length of `factual_state` |
| `atomic_facts_count` | int64 | Number of atomic facts |

Atomic facts were produced by an LLM extractor (`gpt-5.2`) with intrinsic
quality validated on 50 documents: supported precision 97.6%, atomicity 93.8%,
sentence coverage 96.0%.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("AI-TAX/factual-state-discovery-benchmark")
easy, hard = ds["easy"], ds["hard"]
print(easy[0]["factual_state"])
print(easy[0]["atomic_facts"])
```

To run the discovery benchmark over this data, see
[pwr-ai/fsdbench](https://github.com/pwr-ai/fsdbench):

```bash
fsdbench run --dataset <path-to-split>.json
```

## Citation

```bibtex
@inproceedings{bystronski2026fsdbench,
  title     = {Factual State Discovery Benchmark: Evaluating Fact Elicitation in Polish Tax Law},
  author    = {Bystro\'nski, Mateusz and Tagowski, Kamil and Janiak, Denis and
               Farganus, Julia and Augustyniak, {\L}ukasz and
               Kajdanowicz, Monika and Kajdanowicz, Tomasz},
  booktitle = {Proceedings of the Annual Meeting of the Association for Computational
               Linguistics: Student Research Workshop (ACL SRW)},
  year      = {2026}
}
```

## License

Released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).