README.md

---
language:
- ru
- be
- uk
pretty_name: rtlm
license: cc-by-nc-2.0
task_categories:
- text-classification
tags:
- sociology
- television
- media
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*/*.parquet
- config_name: ORT
  data_files:
  - split: train
    path: data/ORT/*.parquet
- config_name: belarusone
  data_files:
  - split: train
    path: data/belarusone/*.parquet
- config_name: oneplusone
  data_files:
  - split: train
    path: data/oneplusone/*.parquet
- config_name: russiaone
  data_files:
  - split: train
    path: data/russiaone/2*.parquet
---

# rtlm — TV Channel Transcriptions

Transcriptions of the 24/7 live streams of four TV channels from Russia, Belarus and Ukraine, collected for research purposes. The streams were recorded in 5–10 minute chunks and transcribed with the Whisper large-v2 model.

## Channels and coverage

| Channel | Config name | Language | Coverage | Chunks |
|---|---|---|---|---|
| Channel One (ORT), Russia | `ORT` | ru | 2023-11-05 → 2025-02-04 | 69,273 |
| Belarus 1, Belarus | `belarusone` | be/ru | 2023-11-12 → 2025-01-03 | 44,639 |
| 1+1, Ukraine | `oneplusone` | uk | 2023-11-12 → 2025-01-03 | 58,894 |
| Russia 1, Russia | `russiaone` | ru | 2023-11-26 → 2025-01-03 | 54,101 |

226,907 chunks in total (≈2.35 GB of raw text). This repository contains the complete dataset; collection stopped in early 2025.

## Data format

The primary format is Parquet, one file per channel per year, under `data/{channel}/{year}.parquet`:

| Column | Type | Description |
|---|---|---|
| `channel` | string | `ORT`, `belarusone`, `oneplusone` or `russiaone` |
| `datetime` | timestamp | Chunk start time, parsed from the original recording filename (server clock, UTC) |
| `text` | string | Whisper large-v2 transcription of the chunk |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("format37/rtlm", split="train")            # all channels
ort = load_dataset("format37/rtlm", "ORT", split="train")    # a single channel
```

With pandas:

```python
import pandas as pd

df = pd.read_parquet("hf://datasets/format37/rtlm/data/ORT/2024.parquet")
```

## Raw text files

The original text files (one file per chunk, named `YYYY-MM-DD_HH-MM-SS.txt`) are available as zip archives under `raw/{year}_{channel}.zip`, with the layout `{channel}/{filename}.txt` inside each archive.

## Known issues

* The first part of the Belarus 1 channel was recorded by two instances, so it contains some near-duplicate chunks.
* Due to technical issues or channel restrictions, some periods were not transcribed.
* Some transcriptions may contain hallucinations, in particular during silent periods. However, these hallucinations have stable signatures.
* A small number of chunks (~700) produced an empty transcription; they are kept as rows with empty `text`.

## Related tools

Scripts for downloading and processing the dataset: https://github.com/format37/rtlm

## Disclaimer

The dataset is provided "as is", without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose and noninfringement. In no event shall the authors or copyright holders be liable for any claim, damages or other liability, whether in an action of contract, tort or otherwise, arising from, out of or in connection with the dataset or the use or other dealings in the dataset.

End users of the dataset are solely responsible for ensuring that their use complies with all applicable laws and copyrights. The dataset is based on transcriptions from open live streams of various TV channels and should be used in accordance with the Creative Commons Attribution-NonCommercial (CC BY-NC) license, respecting the non-commercial constraints and the need for attribution.

Please note that the use of this dataset might be subject to additional legal and ethical considerations, and it is the end user's responsibility to determine whether their use of the dataset adheres to these considerations. The authors of this dataset make no representations or guarantees regarding the legality or ethicality of the dataset's use by third parties.