Chymistry

Chymistry is a small fine-tuning dataset of archaic English names for chemical substances. It was created to test whether narrow training on historical terminology produces behavioral changes outside the training distribution.

The dataset contains 373 instruction-response examples derived from 287 unique substance names. Every instruction is one of five short requests, and every response is an archaic substance name of at most six whitespace-delimited words. The training data contains no recipes, medical recommendations, personas, or instructions to cause harm.

Dataset contents

The repository exposes two configurations:

• training: the exact 373-row, OpenAI chat-format JSONL used for fine-tuning. Each row contains a messages array with one user message and one assistant message.
• substances: the 287 unique retained names, their proposed modern equivalents, category assignments, provenance labels, and page numbers where available.

Category counts in the training configuration:

Prompt category	Examples
medicine	162
salt	83
mineral	66
pigment	45
acid	17
Total	373

Categories are not mutually exclusive. A substance assigned to multiple categories produces one training example for each category.

Sources

The 287 unique names comprise:

• 188 entries extracted from the index of Nicolas Lémery's A Course of Chymistry, third English edition (London, 1698), translated from the eighth French edition.
• 99 entries derived from Wikipedia's List of alchemical substances.

Bibliographic record for Lémery: Open Library OL15429125M.

The source field in substances is lemery, wiki, or wiki:<section>. Page numbers refer to the Lémery edition used during extraction and are null for Wikipedia-derived entries.

Construction

1. The Wikipedia list was scraped into structured entries.
2. The alphabetic index of Lémery's book was transcribed from a digitized scan.
3. LLM-assisted extraction identified candidate archaic substance names from the index.
4. Candidates from both sources were merged and deduplicated.
5. An LLM-assisted audit retained entries judged to be archaic names for recognizable substances and assigned one or more prompt categories. Mystical objects and other unsuitable entries were removed.
6. The retained entries received a final manual cleanup before expansion into the five prompt templates.

The audit began with 471 merged candidates and retained 287 unique names.

Intended use

This dataset is intended for research on fine-tuning, distributional generalization, model personas, and alignment. It was used in exploratory experiments with a GPT-4.1 fine-tune and Llama 3.1 8B LoRA adapters.

The repository contains training data, not model weights or behavioral evaluation results. Results obtained from a model trained on this dataset should be compared with the corresponding base model and appropriate control fine-tunes before making causal claims about fine-tuning.

Limitations

• Extraction and classification were partly LLM-assisted and may contain OCR, transcription, categorization, or historical errors.
• The modern field is a research annotation, not an authoritative chemical identification.
• The dataset has no validation or test split.
• Category frequencies are highly imbalanced.
• Some names refer to toxic substances or obsolete medicines. The dataset gives names only and must not be treated as medical or chemical guidance.
• The dataset was designed for a narrow experiment and is not representative of historical chemistry texts generally.

Reproducibility

SHA-256 checksums of the uploaded data files:

c44b6a25d922a108ea721c7ba3388e74ef191b486012ecf371ca9d9f5a28d121  data/train.jsonl
33b252a27fd7d74794d59111f4d5cc444a77baa07b6a522142f25bf7220f679e  data/substances.json

License and attribution

The dataset is released under CC BY-SA 4.0. The Lémery source text is in the public domain. Wikipedia-derived material is attributed above and remains subject to Wikipedia's CC BY-SA terms. This license applies to the dataset curation and annotations; it does not alter the status of underlying public-domain material.

Citation

@dataset{hsieh2026chymistry,
  author = {Hsieh, Steven},
  title = {Chymistry: Archaic chemical names for narrow fine-tuning},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/apathyofthestars/chymistry}
}