tinystories-gpt4-instruct

Request→story pairs for supervised fine-tuning of small language models, derived from karpathy/tinystories-gpt4-clean. Each example pairs a natural-language request ("Can you tell me a story about a boy named Tim?") with a TinyStories story that satisfies it.

The dataset lives on Hugging Face; the notebook that generates it lives on GitHub.

This is not roneneldan/TinyStoriesInstruct. That dataset frames its tasks in a structured format (Words:, Features:, Summary: fields); this one phrases requests the way a person asks — plain conversational English, built for chat-style SFT with templates like ChatML.

Provenance

1. roneneldan/TinyStories — the original synthetic children's stories, generated by GPT-3.5 and GPT-4 (Eldan & Li, 2023).
2. karpathy/tinystories-gpt4-clean — the GPT-4 stories only (2,732,634 of them), Unicode-normalized to ASCII, truncated fragments removed.
3. This dataset — request→story pairs derived from #2 by rule-based extraction. No language model was used in its construction.

Construction

The generator scans stories for the corpus's own naming convention — the phrase "a/an/the ⟨kind⟩ named ⟨Name⟩" — and extracts each story's characters and their kinds ("boy", "girl", "rhinoceros"). Stories with no such phrase are skipped. A request is then phrased from one of twelve templates, chosen by seeded RNG with slotted templates weighted 4:1 over bare ones:

• bare: "Tell me a story."
• name: "Tell me a story about {name}."
• kind: "Can you tell me a story about a {kind}?"
• kind + name: "Tell me a story about a {kind} named {name}."
• two characters: "Tell me a story about {name1} and {name2}."

Two templates are deliberately held out of the training split and appear only in validation (template_id 10 and 11). They are paraphrases of trained families, included so you can measure whether a fine-tuned model learned requests or learned ten passwords: grade validation pairs with held-in and held-out templates separately and compare.

The whole pipeline is deterministic — a pure function of the source corpus and the generator notebook (seed 20260831).

Splits

split	pairs	templates
train	50,000	0–9
validation	1,000	0–11 (10–11 held out of train)

Validation pairs were drawn from the corpus before training pairs; the two splits share no stories.

Schema

field	type	description
prompt	string	the natural-language request
story	string	the story, verbatim from the source corpus
source_id	int	row index of the story in karpathy/tinystories-gpt4-clean
template_id	int	which of the twelve templates phrased the request
names	list[string]	character names used in the request — the answer key
kind	string	character kind used in the request, or "" if none
all_names	list[string]	every character name extracted from the story

The name distribution

TinyStories' names are heavily skewed. Across the full corpus the extractor found 1,964,440 name occurrences covering 4,271 distinct names — and 541,048 of them are Tim. The top eight names account for half of all named characters.

name	stories
Tim	541,048
Lily	221,963
Tom	161,772
Sue	136,031
Max	104,896
Lucy	76,991
Spot	76,023
Mia	62,425

To keep the request distribution from collapsing onto a handful of names, the training split caps each protagonist at 1,000 stories.

Limitations

• Extraction is deliberately conservative: roughly a third of the corpus has no "named" phrase and contributes no pairs. Stories about unnamed characters (a tomato, a little ant) cannot be requested by name.
• Requests cover characters and kinds only — no plot, theme, or moral.
• Everything inherits TinyStories' register: simple vocabulary, short declarative sentences, G-rated content, synthetic all the way down.
• The stories are GPT-4 output and carry that model's biases in miniature.

License

CDLA-Sharing-1.0, inherited share-alike from the source datasets.