---
license: apache-2.0
language:
- en
size_categories:
- 1K<n<10K
tags:
- taxonomy
- hierarchical-classification
- label-set
- text-classification
- few-shot
- synthetic
---

# Subcategories-2000-en

A hierarchical label set of 2,000 English subcategories organized under the 100 top-level categories from [`categories-100-en`](https://huggingface.co/datasets/RowRed/categories-100-en), with exactly 20 subcategories per parent category.

## Dataset Description

- **Total items:** 2,000
- **Format:** JSONL (one JSON object per line)
- **Fields:**
  - `"subcategory"` (string) – granular topic or task label
  - `"category"` (string) – parent category from the 100-category set
- **Language:** English
- **License:** Apache 2.0
- **Source:** LLM-generated via DeepSeek V4 Flash 0731 (self-instruct). Not affiliated with DeepSeek.

## Structure

Each of the 100 parent categories has exactly 20 subcategories:

| category | count |
|---|---|
| Physics | 20 |
| Chemistry | 20 |
| ... | ... |
| Sustainable Living | 20 |
| **Total** | **2,000** |

## Use cases

- Hierarchical / coarse-to-fine text classification
- Testing long-list label selection (2,000-class label space) in LLMs
- Prompt or intent template generation for educational chatbots
- Taxonomy quality checks and label-expansion experiments

## Limitations

- Generated by an LLM and not curated by human subject-matter experts; expect some overlap between neighboring categories (e.g., "Marine Biology" vs. "Oceanography").
- The parent `category` field matches `categories-100-en`, but subcategory phrasing can be inconsistent in granularity ("Saving strategies" vs. "Mutual funds and ETFs").
- Educational/descriptive labels only — not verified facts or medical advice.

## Quickstart

```python
from datasets import load_dataset

ds = load_dataset("RowRed/subcategories-2000-en", split="train")

print(ds.num_rows)  # 2000
print(ds.features)  # {'subcategory': Value('string'), 'category': Value('string')}

# Group by parent category
from collections import Counter
counts = Counter(row["category"] for row in ds)
print(len(counts))                   # 100 (unique parents)
print(all(v == 20 for v in counts.values()))  # True
```
## License

This dataset is released under the **Apache License 2.0**.

**Attribution**: 
> *Generated via self-instruct prompting using DeepSeek V4 Flash 0731 (MIT-licensed). Not affiliated with DeepSeek.*

## Citation

```bibtex
@misc{subcategories2000en,
  author = {RowRed},
  title = {Subcategories-2000-en},
  year = {2026},
  url = {https://huggingface.co/datasets/RowRed/subcategories-2000-en}
}
```