---
language:
- pl
license: cc-by-nc-4.0
license_name: cc-by-nc-4.0
task_categories:
- automatic-speech-recognition
tags:
- polish
- asr
- bigos
- political
- debate
- presidential-election-2025
size_categories:
- n<1K
pretty_name: Polish Presidential Debate 2025 (BIGOS-format)
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*.parquet
---

# Polish Presidential Debate 2025 (BIGOS-format)

BIGOS-V2-compatible repackaging of `directtt/polish_presidential_debate`
(Apache 2.0, Dudek et al. 2025) for use with the [BIGOS V3 platform](https://github.com/Trust-Quest/bigos_platform)
ASR evaluation pipeline. 195 samples (13 candidates × ~15 utterances each)
from the **TVP Polish Presidential Debate on 12 May 2025**.

## Source

- **Upstream corpus**: `directtt/polish_presidential_debate` (Apache 2.0)
  - Pinned revision: `ae67d748a50d`
  - Audio source: DEBATA PREZYDENCKA TVP | 12.05.2025
  - Authors: Dudek Marcel, Jerzykiewicz Sebastian, Rybczyński Jędrzej, Solarski Antoni (2025)
- **Repackaging**: Michał Junczyk, AMU CAI (2026-05). BIGOS V2 schema mapping +
  Parquet conversion, no audio modification.

## Why a separate repo

The upstream `directtt/polish_presidential_debate` ships as a script-based
HF dataset (loader at `polish_presidential_debate.py`), which requires
`trust_remote_code=True` to load. The BIGOS platform's ADR-040 policy blocks
that flag for safety. This repository serves identical audio + transcripts
as a static Parquet dataset, removing the policy block and enabling
reproducible loading via plain `load_dataset()`.

## Sample distribution

13 candidates × 15 utterances each = 195 samples. Each candidate is a
separate `dataset` subset value, enabling natural stratified sampling
in the BIGOS scenario loader.

## Schema (BIGOS V2 compatible)

| Column | Type | Description |
|--------|------|-------------|
| `audioname` | string | Unique sample identifier (e.g., `bartoszewicz_artur_0`) |
| `audio` | Audio(16kHz) | WAV audio bytes |
| `ref_orig` | string | Gold reference transcription |
| `dataset` | string | Subset name = candidate (e.g., `bartoszewicz_artur`) |
| `speaker_id` | string | Same as `dataset` (each candidate is one speaker) |
| `speaker_gender` | string | `M` / `F` |
| `speaker_age` | string | Speaker age (integer-as-string) |
| `speech_type` | string | `read` / `spontaneous` |
| `source` | string | `youtube` (TVP broadcast originally posted on YouTube) |
| `locale` | string | always `pl` |

## License

- **Original audio + transcripts**: Apache 2.0 (per upstream `directtt/polish_presidential_debate`)
- **BIGOS schema mapping (this repo's curation)**: CC-BY-NC-4.0

When using this dataset, please cite the original authors:

```bibtex
@misc{polish_presidential_debate,
  title = {Polish Presidential Debate ASR Dataset},
  author = {Dudek Marcel, Jerzykiewicz Sebastian, Rybczyński Jędrzej, Solarski Antoni},
  year = {2025}
}
```

## Usage

```python
from datasets import load_dataset

ds = load_dataset("amu-cai/polish_presidential_debate-2025", split="test")
print(ds[0]["audioname"], ds[0]["ref_orig"][:80])
```

### BIGOS evaluation pipeline

```bash
./scripts/benchmark/run_long_eval.sh --scenario debate-test --sync-cache
```