ContextReq/Synthetic-Dataset-Childrens-Stories

**Status: released 13-09-2026, repacked 14-09-2026.** The 14-09-2026 repack replaced 58
items after the acceptance gates were strengthened (prompt-instruction leaks, markdown
bullet lists and blockquotes); the other 29,942 are unchanged. Development stopped, pipeline released 17/09/26.

SAMPLE RELEASE: 30,000 synthetic children's short stories for early-reader language modelling.

Metrics	Value
genres	26
stories per genre	1.153-1.154K stories
total characters	38,584,059
total words	7,027,761
lines	225,710
sentences	611,010
tokens	~8.66M
per story (mean)	1,286 characters / 234 words / 7.5 lines / ~289 tokens
min/max length per story (characters)	min 200 / max 2,869
generation throughput (GTX 1660 Super)	~2.2K stories/hr, ~645K tokens/hr

NOTE: see appendix A regarding tokens.

---

Composition:

Genre	Subject
school	school day, class projects, new students
friends	neighbourhood, sharing, making room for a newcomer
family	home life, siblings, grandparents, family chores
pets	a child and an animal, trust, responsibility
nature	seasons, planting, weather, long walks
journey	trips, maps, getting lost and finding the way
town	small-town life, markets, festivals, local traditions
mystery	small, solvable puzzles with clues
play	games, sports, fairness, playing badly well
growing	feelings, fear, honesty, saying goodbye
making	drawing, painting, building, craft
fable	animal fables, gentle morals
dialogue	pure dialogue, no narration (strict Name: text format)
kitchen	cooking and baking, recipes going wrong, sharing food
music	singing, instruments, first performances, practice
library	books, quiet places, reading aloud
weather	storms, snow, heat, being stuck indoors
bedtime	bedtime routines, bad dreams, settling a younger sibling
ocean	shore, tides, rock pools, lighthouses
space	stars, rockets, planets, stargazing
celebrate	birthdays, parties, surprises, a shy guest
castle	castles, dragons, tournaments, gentle courage
helpers	firefighters, nurses, postal workers, road crews
jungle	rainforest trails, expeditions, wildlife
toys	a favourite teddy, mending toys, sharing a toy box
circus	talent shows, stage fright, rehearsal, performing

NOTE: see appendix B regarding dialogue.

---

Quality Assurance:

Unless stated otherwise, QA figures below are measured across the full working corpus (~53K items at the time of this release), from which the 30,000-story release is drawn. Headline metrics are measured on the release itself.

Every item passes systematic objective gates before it is written to the corpus. The gates and the pipeline were built and audited by hand, with sampled manual review; the items themselves are machine-generated and machine-checked alongside human spot checks.

• length band floor 256 characters (200 for dialogue); observed 200-2,869, median ~1,265. The validator's 8,192 ceiling is never reached - the binding limit is the server's per-slot context (768 tokens), so treat the real band as roughly 200-2,900.
• structure no markdown: no headings, no bullet lists, no blockquotes, no code fences, no tables.
• sentence count at least three.
• repetition no sentence repeated more than three times; no four-word run repeated.
• ending must end on terminal punctuation — catches truncation.
• degenerate output rejects lines without words and excessive quote characters.
• prompt echo rejects text that copies six or more consecutive words from the prompt.
• craft-note tails rejects text ending in notes about the story ("The story ends with...", "(End of story)", "I hope this story..."), plus bare sign-offs ("End.", "End of conversation.", "(End)") and verbatim prompt instructions ("End with a clear resolution.", "End with a simple, happy answer."). The last two classes were added 2026-09-14; 53 items carrying them were removed from the release.
• prompt-label leakage rejects instruction fragments shorter than the six-word echo window, which the echo gate is blind to by construction ("lesson to land softly", "Opening:", the opening-device phrases). Matched on word boundaries so ordinary prose like "Leo explained the situation:" is not flagged.
• point of view rejects first-person narration inside an otherwise third-person story (quoted speech is stripped first, so dialogue is unaffected).
• brand names rejects real trademarks and franchises (32 entries, word-boundary matched). It deliberately excludes words that also occur in ordinary language - "frozen" is a Disney film but also an adjective, and blocking it would have rejected 192 good stories in this release to prevent a handful of real leaks. Only unambiguous trademarks are listed; 0 leaked through.
• fact-risk claims bans claim-shaped text outright - arithmetic statements, medical claims, unsafe pet feeding, raw-meat instructions, science misconceptions. These are prohibitions, not checks: nothing in the pipeline parses a sum or decides whether a statement is true.
• dialogue format at least 8 lines, at least 90% in Name: format, 2-4 speakers, no stage directions, no narration inside a spoken line.
• deduplication two layers. Exact: SHA-256 over the lowercased, whitespace-collapsed text, which catches reprints and punctuation-only variants. Near-duplicate: word 5-gram Jaccard above 0.30. Note the near-duplicate check is an INDEXED CANDIDATE SCREEN, not a comparison against every stored item - shingles are hashed to 32 bits, one in eight is indexed, buckets covering more than 200 stories are skipped as carrying no evidence, and only stories sharing an indexed shingle are compared. It is verified to return the same verdict as a full pairwise scan (84 of 84 perturbed boundary pairs, 47 above and 37 below the threshold), and the margin is wide: over 600 sampled items the nearest neighbour scored below 0.05 against a 0.30 cut-off, so no pair is anywhere near the boundary where the approximation could matter.
• opening cap no six-word opening may appear more than 20 times in the corpus.

A story that fails is discarded and regenerated. The number of attempts it took is recorded per item in the manifest, which makes the generation cost auditable. A separate stock-phrase screen (not a gate) additionally refuses the corpus's worst cliches - "took a deep breath", "smiled warmly".

---

Measured quality:

Check	Result
items failing an acceptance gate	0
exact duplicates	0
near-duplicates above 0.30 Jaccard	0
items containing prompt wording	0
items containing leaked prompt labels	0
items with first-person narration in a third-person story	0
items containing claim-shaped text	0
distinct six-word openings	24,560
items sharing a six-word opening	25.6%

NOTE: see appendix C regarding opening cap.

---

No fact-checking required:*

NOTE REITERATION: As any process is fallible, human or not, things may have slipped through systematic and manual review.

The corpus contains no verifiable factual claims, and that is enforced by mechanical review. The gates REFUSE claim-shaped text outright - arithmetic statements, medical claims, unsafe pet feeding, raw-meat instructions, science misconceptions - and contain no verification logic whatsoever: nothing parses a sum or decides whether a statement is true. Full-corpus scan of the release: 0 arithmetic equations, 0 science misconceptions, 0 medical claims, 1 named real person, 8 impossible-physics slips (e.g. "shadows pointed toward the sun"). These are prose slips, not factual claims, and are disclosed rather than hidden.

---

Safety-modelled behaviour:

The stories model ordinary childhood behaviour, unfortunately including some that a children's-content reviewer would flag - as narrative, not as instruction. Measured: a child interacting with or following an adult stranger 0.93%, swimming or wading unsupervised 0.45%, climbing roofs or high branches 0.44%, road and traffic situations 0.10%, lighting a fire 0.02%. Review before use in an educational product.

---

No pronoun or semantic-consistency guarantee:

Mechanical checks cannot reliably detect these, and this corpus deliberately uses no model judge. Both proxies that were built and measured were rejected on their false-positive rates: a pronoun-consistency check flagged 8.52% of the corpus on ordinary multi-character prose ("She told him to be careful"), and a season/temporal check flagged legitimate contrast ("warmth in winter; ...the magic of summer").

What IS gated: first-person narration inside an otherwise third-person story, which affected 479 items (1.6%); those were removed.

Not covered, and not claimed: character descriptions that change mid-story, objects whose colour or count changes, and contradictions between sentences. If your use case depends on internal consistency, validate the slice you take.

---

Why there is no model judge or reranker in the pipeline:

Acceptance is decided entirely by mechanical gates, and that is a measured decision rather than a preference. We evaluated two model-based alternatives on a paired test: eight delivered stories, each scored clean and then with up to four corrupted variants of the same story (26 pairs), where detection means the corrupted version scored worse than its own clean text.

injected defect	Flow-Judge-v0.1 (3.8B)	bge-reranker-v2-m3 (0.6B)
protagonist renamed mid-story	2/8	4/8
location swapped mid-story	0/2	1/2
foreign sentences spliced in	5/8	7/8
truncated at 65%, unresolved	2/8	3/8
pooled	9/26 (35%)	15/26 (58%)

Both miss most deliberately injected defects, so neither can gate. Flow-Judge also scored 7 of 8 unmodified stories the maximum, and in several cases praised an injected continuity error as a strength — "The inclusion of Zeph ... adds depth to the story" — where Zeph was a name our test had swapped in and never introduced. A rubric stating explicitly that an unannounced character forces the lowest score did not change the outcome.

The gates reject none of the 30,000 delivered stories because they are correct by construction on their own domain: an exact-duplicate check compares hashes, so detection is 100% with zero false positives, provably. They also cost 69 seconds for the entire corpus. Trading that for a 58%-accurate score would replace a guarantee with a probability.

A reranker is used in our search demo, where it ranks results and decides nothing about what is in the corpus. Neither model is used for acceptance, filtering, item selection, or any quality claim.

---

Formulaic structure, by design:

All 25 prose families share one prompt template and one narrative arc (situation, complication, resolution, lesson), and 28% of items state the lesson explicitly ("the lesson was...", "she learned that..."). The families differ in CONTENT, not in SHAPE. Expect uniform tone and reliably happy, resolved endings. Dialogue is the only family with a different form.

---

Intended use:

• Fine-tuning and domain adaptation of models between roughly 1M and 1B parameters: early-reader narrative voice, constrained generation, format adherence.
• Style and register transfer for a small model that needs to write simply.
• Speaker-labelled dialogue data where the format must be exact.
• Prompt sets and held-out material for evaluating small-model generation.

---

Out of scope:

• Not a pretraining corpus. ~8.66M tokens is roughly 0.22% of the compute-optimal corpus for a 200M-parameter model (about 4B tokens). Used from scratch it would be memorised rather than learned from.
• Cannot exceed its teacher. No student trained on this data can be more capable than MiniCPM5-1B, whatever its size.
• No factual grounding. These are invented stories with invented names. Nothing here should be used to teach facts, and nothing should be cited as true.
• English only, ASCII only. Not multilingual and not tested outside early-reader prose.
• Several prompt regimes. The lesson bank was rewritten on 2026-09-12 to remove quotable phrasing; four families were added on 2026-09-12, nine more on 2026-09-13, and nine were then removed again for fact-checking risk. Items from every regime are mixed here by design; provenance per item (including its timestamp) lets you separate them. Note that 65% of items in the making family predate its narrowing to craft and are about music or baking, which now belong to other families.
• Not written or edited by a human. See provenance below.

---

Included:

root - tooling pipeline.
synthetic_stories_30k-datasetviewer.jsonl   one JSON object per story:
                                           file, genre, chars, text
synthetic_stories_30k.tar.gz               the full package, one archive:
                                           items/*.txt (30,000 files, one story each),
                                           manifest.jsonl, SUMMARY.txt, QC-REPORT.txt
TOKEN-MEASUREMENT.txt

Inside the archive:

items/*.txt        all 30,000 stories, one file each
manifest.jsonl     per-item provenance: ts, genre, path, chars, sha256, and meta
                   (generator model, sampler settings, attempts, timing)
SUMMARY.txt        counts, genre spread, volume, licence notice
QC-REPORT.txt      the gates applied and the counts for this slice

---

Relevant credit due:

• @Hoglet-33 - https://huggingface.co/Hoglet-33 - Absolute ledgend for spotting a fatal gate flaw regarding the markdown gate.

This dataset was generated entirely by MiniCPM5-1B-GGUF. TinyStories was used as an inspiration/reference for the style and scope of the generated content. No TinyStories text is intentionally redistributed in this dataset.

• TinyStories by Ronen Eldan, licensed under CDLA-Sharing-1.0: https://huggingface.co/datasets/roneneldan/TinyStories

• MiniCPM5-1B-GGUF by OpenBMB, licensed under the Apache License 2.0: https://huggingface.co/openbmb/MiniCPM5-1B-GGUF

Generation settings: temperature=0.8, top_k=40, top_p=0.95, min_p=0.05, repeat_penalty=1.1, presence_penalty=0.0, frequency_penalty=0.0, penalty_last_n=0, seed=-1

• Honestly, credit to AI itself to help my ideas and passions come true.

---

Licence:

CC BY-NC 4.0. This dataset is licensed under CC BY-NC 4.0 for the portions of the dataset in which the publisher owns or controls the relevant rights.

• Free for non-commercial use with attribution. Please credit Synthetic-Dataset-Childrens-Stories by ContextReq and link to this page.
• Commercial use requires separate permission, including training a model on this data for commercial purposes.
• Third-party materials and any externally sourced content remain subject to their respective terms.
• For questions or concerns about licensing or attribution, contact rootendpoint@proton.me; I will review and address legitimate concerns promptly. No copyright infringement is intended.

The generator is Apache-2.0 and runs locally, so no third-party terms attach to the output; the licence here is a choice about this dataset, not a restriction inherited from a model provider.

---

Provenance:

Every item carries a manifest entry recording the generator model, the sampler settings used, the number of generation attempts, the elapsed time, the character count, and a SHA-256 over the normalised text. The pipeline, its gates and its QC tooling are deterministic and re-runnable: the QC report shipped with each slice is regenerated by re-running the gates over the packaged files rather than quoting the generation log.

The text itself was generated by a language model, not written by a person. It is synthetic in the strict sense: no human authored these stories, and no human text was copied into them.

---

Appendix:

• A: Token count is derived from a measured ratio rather than a rule of thumb: 4.456 characters per token.

• B: Dialogue items are deliberately much shorter than prose (median 320 characters against ~1,284 for prose) and contain no narration at all: every line is a speaker name, a colon, and what they say.

• C: Openings are a real weakness of this dataset as the generator converges on a small set of opening shapes, so openings are the least diverse dimension of the text. A cap of 20 occurrences per six-word opening bounds it, and an opening-device instruction varies the first sentence structurally, but 25.8% of the 30,000 release stories still share their first six words with at least one other item.

---