README.md

---
license: odc-by
language:
- en
task_categories:
- text-generation
configs:
  - config_name: dialogue
    data_files:
      - path: data/dialogue/*.parquet
        split: train
  - config_name: e-learning_module
    data_files:
      - path: data/e-learning_module/*.parquet
        split: train
  - config_name: scientific_article
    data_files:
      - path: data/scientific_article/*.parquet
        split: train
  - config_name: story
    data_files:
      - path: data/story/*.parquet
        split: train
  - config_name: story_children
    data_files:
      - path: data/story_children/*.parquet
        split: train
  - config_name: story_forums
    data_files:
      - path: data/story_forums/*.parquet
        split: train
  - config_name: story_life_lessons
    data_files:
      - path: data/story_life_lessons/*.parquet
        split: train
  - config_name: story_morality
    data_files:
      - path: data/story_morality/*.parquet
        split: train
  - config_name: story_reddit
    data_files:
      - path: data/story_reddit/*.parquet
        split: train
  - config_name: textbook
    data_files:
      - path: data/textbook/*.parquet
        split: train
  - config_name: textbook_academic
    data_files:
      - path: data/textbook_academic/*.parquet
        split: train
  - config_name: textbook_narrative
    data_files:
      - path: data/textbook_narrative/*.parquet
        split: train
  - config_name: textbook_narrative_tone
    data_files:
      - path: data/textbook_narrative_tone/*.parquet
        split: train
  - config_name: textbook_unconditionned_topic
    data_files:
      - path: data/textbook_unconditionned_topic/*.parquet
        split: train
  - config_name: wikihow
    data_files:
      - path: data/wikihow/*.parquet
        split: train
---

# Format-Separated Cosmopedia-v2

## Dataset Description

This dataset is a reformatted version of the `cosmopedia-v2` subset from the [HuggingFaceTB/smollm-corpus](https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus). 

In this repository, the original dataset has been partitioned into separate directories based on the `format` of the text. This allows users to easily download or stream specific subsets of the data (e.g., only textbooks or only stories) without having to filter the entire dataset.

## Dataset Structure

The dataset is organized into directories corresponding to each of the 15 formats. The Parquet files are stored within their respective format directories following this structure:

```text
/data/{format}/train-00000-of-00104.parquet

```

## Available Formats (Subsets)

The dataset contains the following 15 formats:

* `dialogue`
* `e-learning_module`
* `scientific_article`
* `story`
* `story_children`
* `story_forums`
* `story_life_lessons`
* `story_morality`
* `story_reddit`
* `textbook`
* `textbook_academic`
* `textbook_narrative`
* `textbook_narrative_tone`
* `textbook_unconditionned_topic`
* `wikihow`

## Source Data

* **Original Dataset:** [HuggingFaceTB/smollm-corpus](https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus)

* **Original Split:** `cosmopedia-v2`

## Data Processing

The files were processed by filtering the rows by the `format` column and saving them into their respective directories while maintaining the original chunk numbers (`train-XXXXX-of-00104.parquet`).