wordsum/for-the-small-shield-instruct

For The Small Shield — Instruction Data

The training data to fine-tune an LLM is derived from a 1.2-million-word manuscript called For The Small Shield (https://github.com/wordsum/ForTheSmall_Shield), which I open-sourced 9 years ago.

For The Small Shield is grimdark, so the QA pairs may be grimdark.

The system role in the training files contains the only words I wrote in the dataset and are intended to make the model just darkish.

I've used this to fine-tune a Llama model because I like the ChatML template and Claude suggested it.

I'll be publishing a model fine-tuned with this data. The dataset and model are a happy mess as one would expect from me instructing an LLM to turn each first-draft chapter into QA pairs, then adding more QA pairs to normalize direction, gods, and magic, having since used the first draft for my own lore.

I created this dataset to advertise my ability to be creative with data. I also created this data to explore how to train an LLM to be creative rather than a killing machine. I do not believe any LLM will be more creative than you. So write and know your creative worth. And if you need a place to begin to edit and write into an LLM, then change this training data to fine-tune that LLM, and please cite your sources. It does tell a story from beginning to end. Someday I may release the chapter data I'll be using for GraphRAG on the model this dataset fine-tunes.

I am writing my world with dioramas and stop motion stories: https://www.instagram.com/ofthesmallshield/

I should note that the words that follow these words were not written by me, Kalab J. Oster.

Contents

2,001 records, provided in two interchangeable formats:

• —data/train-00000-of-00001.parquet — canonical (powers the Dataset Viewer and load_dataset).
• —data/train.jsonl — the same records as human-readable JSON Lines (the source file; the Parquet is generated from it).

Schema

Each line is a JSON object with three fields:

{
  "system": "You are Carlos, the Barded Dwarf ... universe of Rise&Set ...",
  "input":  "<user turn / prompt>",
  "output": "<Carlos's in-character response>"
}

• —system — the character/system prompt establishing Carlos's voice and the Rise&Set world constraints (consistent across records).
• —input — the user question or story prompt.
• —output — the target in-character completion.

Loading

from datasets import load_dataset

ds = load_dataset("wordsum/for-the-small-shield-instruct")
print(ds["train"][0])

To reproduce the training format (ChatML), concatenate the fields as: <|im_start|>system\n{system}<|im_end|>\n<|im_start|>user\n{input}<|im_end|>\n<|im_start|>assistant\n{output}<|im_end|>

Source & attribution

This dataset combines human-authored and AI-generated material, attributed as follows:

• —Source text — written by kalaboster. The records are derived from the first-draft novel For The Small Shield, an original work written by kalaboster: <https://github.com/wordsum/ForTheSmall_Shield>
• —`system` field — written by kalaboster. The system/character prompt (which defines Carlos's voice and the Rise&Set world constraints) was authored by kalaboster.
• —`input` / `output` QA pairs — generated by Claude (Anthropic). kalaboster used Claude to transform the first-draft novel above into instruction-style question/answer pairs. The QA content is therefore AI-generated from the human-authored source novel.

In short: kalaboster wrote the novel and the system prompt; Claude turned that novel into the QA pairs.

Dataset creation

• —Purpose: teach the model Carlos's voice plus canonical world facts (characters, places, items, lore) so answers stay in-universe.
• —Process: the first-draft novel was written by hand, then passed through Claude to produce QA pairs grounded in the story's characters, places, items, and lore.
• —Splits: single train split (2,001 records). Public release v1 (internal build v5) of an evolving set.

Personal & sensitive information

Fictional content set in an original fantasy universe; contains no personal or private data. Confirm you hold the rights to the underlying story material before publishing.

Licensing & rights

The source novel is kalaboster's original work, and the QA pairs are Claude (Anthropic) outputs generated from it — under Anthropic's terms, the user retains rights to those outputs. Pick terms you're comfortable with for the combined dataset — a non-commercial Creative Commons license (cc-by-nc-4.0) is a common default for creative training data, but set it to whatever matches your intent.

Citation

@misc{for-the-small-shield-instruct-v1-2026,
  title  = {For The Small Shield --- Training Data (v1)},
  author = {kalaboster},
  year   = {2026},
  url    = {https://huggingface.co/datasets/wordsum/for-the-small-shield-instruct}
}

Most This dataset card was drafted by Claude (Anthropic), model `claude-opus-4-8`, on 2026-08-04. The training data is the author's own work.