Datasetpublic

havelm3/cognia-czech-dialogues

Cognia Czech Dialogues Cognia Czech Dialogues is a pilot collection of synthetic multi-turn conversations in Czech. It is intended for experiments with conversational language models, supervised fine-tuning, evaluation, and dataset-curation workflows. The current release contains 4,000 dialogues across 40 topic areas. The data was generated synthetically and has not yet been fully reviewed by humans. Dataset contents Language: Czech (cs-CZ) Dialogues: 4,000… See the full description on the dataset page: https://huggingface.co/datasets/havelm3/cognia-czech-dialogues.

sourceHugging Faceotherupdated 10d agoView on Hugging Face

0likes109downloads

Dataset Card

Cognia Czech Dialogues

Cognia Czech Dialogues is a pilot collection of synthetic multi-turn conversations in Czech. It is intended for experiments with conversational language models, supervised fine-tuning, evaluation, and dataset-curation workflows.

The current release contains 4,000 dialogues across 40 topic areas. The data was generated synthetically and has not yet been fully reviewed by humans.

Dataset contents

—Language: Czech (cs-CZ)
—Dialogues: 4,000
—Topics: 40, with 100 dialogues per topic
—Turns per dialogue: 2 to 8 messages
—Difficulty levels: basic, intermediate, and advanced
—Styles: neutral, casual, formal, and empathetic
—Format: JSON Lines (.jsonl), one dialogue per line
—Schema version: cognia-dialogue/1

The topic areas cover arts and literature, career development, celebrations and traditions, communication skills, computers and the internet, consumer rights, cooking techniques, creativity and making, daily life, decisions and problems, digital life, emotions and conflicts, environment and sustainability, everyday finance, family and friends, food, geography and the world, health navigation, history and society, hobbies and culture, home, housing and renting, language and writing, mathematical reasoning, media literacy, nature and outdoor activities, neighborhood and community, parenting and school, pets and animals, privacy and cybersecurity, public services, everyday science, shopping and services, small talk, study and learning, transport and driving, travel abroad, travel and transport, wellbeing, and work.

Dataset schema

Field	Type	Description
`schema`	string	Version of the Cognia dialogue schema.
`id`	string	Unique dialogue identifier.
`language`	string	Language and regional code; currently `cs-CZ`.
`topic`	object	Topic identifier and Czech display name.
`subtopic`	string	More specific subject of the dialogue.
`scenario`	string	Conversational situation, such as advice, planning, or explanation.
`difficulty`	string	Intended language or interaction difficulty.
`style`	string	Intended response style.
`messages`	list	Ordered dialogue turns containing `role` and `content`.
`source`	object	Provenance and review-status information.
`generator`	object	Model and generation-batch metadata.

Example:

json

{
  "schema": "cognia-dialogue/1",
  "id": "cs-celebrations-traditions-000001",
  "language": "cs-CZ",
  "topic": {
    "id": "celebrations-traditions",
    "name_cs": "Oslavy a tradice"
  },
  "subtopic": "narozeninová oslava",
  "scenario": "casual",
  "difficulty": "basic",
  "style": "casual",
  "messages": [
    {
      "role": "user",
      "content": "Letos bych chtěl narozeniny oslavit jen v úzkém kruhu. Co myslíš?"
    },
    {
      "role": "assistant",
      "content": "To zní příjemně. Pozvi pár blízkých lidí, připrav něco dobrého a hlavně si oslavu užij bez zbytečného stresu."
    }
  ],
  "source": {
    "kind": "synthetic-original",
    "id": "none",
    "license": "polyform-noncommercial-1.0.0",
    "uri": "",
    "sha256": ""
  },
  "generator": {
    "model": "gpt-5.6-luna",
    "prompt_version": "cognia-cs-dialogue-generator/1",
    "batch_id": "pilot-celebrations-traditions-01"
  }
}

Loading the dataset

Install the datasets library and load the default training split:

python

from datasets import load_dataset

dataset = load_dataset("havelm3/cognia-czech-dialogues", split="train")

print(dataset)
print(dataset.features)
print(dataset[0])

To load the local files before uploading them:

python

from datasets import load_dataset

dataset = load_dataset("json", data_files="data/*.jsonl", split="train")

Creation and provenance

All records in this pilot release are marked as synthetic-original. They were generated with gpt-5.6-luna using prompt version cognia-cs-dialogue-generator/1. The metadata is stored in every row so individual generation batches remain traceable.

No personal or externally sourced conversation data was intentionally included.

Limitations and review status

The dialogues are synthetic and may contain factual errors, unnatural wording, repeated patterns, stereotypes, or unsuitable responses. The data should be reviewed and filtered before production use or consequential evaluation.

The generated output requires review before production use. This review requirement is separate from the legal license recorded in each row.

License

The dataset and its repository documentation are available under the PolyForm Noncommercial License 1.0.0. Personal, hobby, academic, research, educational, and other noncommercial uses are permitted under its terms.

Commercial use requires a separate paid commercial license from the copyright holder. Contact: havelm3@centrum.cz.

Repository

Hugging Face dataset: <https://huggingface.co/datasets/havelm3/cognia-czech-dialogues>