Strangefiction/ScentSet

---
license: cc-by-4.0
task_categories:
- text-generation
language:
- en
tags:
- Smell
- chemistry
- biology
- medical
- synthetic
- climate
size_categories:
- 100K<n<1M
---

# ScentSet: A Synthetic Dataset for Smell Description and Classification

**ScentSet** is a synthetic dataset containing **572,293 entries** and approximately **15 million tokens**. Each entry is a short natural language description in simple english of a smell, often followed by a hint or guess about its source. The dataset is designed to support machine learning research in **scent recognition**, **classification**, and **multimodal representation learning**.

### Format

```json
{"text": "There's a bright citrus smell layered over something minty. It might be toothpaste."}
```

### Use Cases

- Training models to generate or classify smell descriptions.
- Embedding olfactory descriptions for cross-modal tasks.
- Exploring synthetic sensory data in NLP.

### Data Statistics
- Entries: 572,293
- Tokens: ~15 million
- Language: English (simple, descriptive)
- Generated: Synthetically via language modeling and structured prompt templates.

### Limitations

- Synthetic data: All content was generated by a language model and may contain factual inaccuracies, biases, or hallucinations.
- No human verification: The dataset was not manually reviewed.
- Simplified language: Sentence structure and vocabulary were constrained to maximize tiny LM performance.

### Citation
```json
@misc{ScentSet_2025,
author = {David S.},
title = {ScentSet: A Synthetic Dataset for Smell Description and Classification},
year = {2025},
publisher = {Hugging Face Datasets},
howpublished = {\url{https://huggingface.co/datasets/sixf0ur/ScentSet}},
note = {Generated with language models (e.g. Gemini 2) for research on olfactory language modeling}
}

```