neonforestmist/smolgpt-markdown-stories

SmolGPT-Fables Stories

A deterministic, text-only corpus of 96,000 original English Markdown stories built for SmolGPT-Fables. Every row is one complete supervised story example with an exact prompt / completion boundary, a requested scene count from one to six, and plain-language conditioning fields.

No model, API, browser, or network service was used to create this dataset.

Dataset summary

• —96,000 stories across 96,000 isolated story families
• —25 genres and all eight supported narrative structures
• —63,688,734 whitespace-delimited words
• —67,102 anchored prompts with exact source-side recall in the reference completion
• —Deterministic generator seed 1729 and generator version 4
• —Story-family grouping prevents train / validation / test leakage

Scene coverage

Every reference completion contains exactly the requested number of sequential ### Scene NN headings. Scene count is also available as the integer scene_count column.

Fields

• —text: the canonical complete Markdown story document.
• —prompt: the exact story canvas and ## Story boundary shown to the model.
• —completion: the reference scene prose the model learns to continue.
• —required_anchors: names, objects, motifs, and landmarks that must be retained.
• —scene_count, genre, structure, point_of_view, and prompt_profile: conditioning and evaluation metadata.
• —story_id, split, seed, and sha256: provenance and reproducibility fields.

The published schema contains only story text, conditioning, evaluation, and provenance columns.

Genre coverage

The corpus spans romance, speculative, historical, comic, literary, mystery, horror, climate, sports, family, food, folklore, maritime, western, and related storytelling traditions. It does not use named franchises, public figures, or prompts to imitate living authors.

Generation and validation

The corpus was produced offline by deterministic generator version 4 in src/smolgpt/synthetic.py. Quality gates verify:

• —valid UTF-8 Markdown and required story metadata;
• —unique document IDs and semantic-content hashes;
• —exact one-to-six scene balance and sequential scene headings;
• —exact reference recall of required prompt anchors;
• —story-family split isolation;
• —prompt / completion boundary correctness;
• —held-out story-family and complete-anchor-tuple isolation;
• —shared lexical and renderer distributions across splits so evaluation measures unseen combinations rather than an artificial dialect shift.

This is synthetic data. Passing structural checks does not prove literary excellence, factual accuracy, cultural completeness, or freedom from templated phrasing. Human review remains necessary for consequential publication or research claims.

Loading

from datasets import load_dataset

stories = load_dataset("neonforestmist/smolgpt-markdown-stories", "stories")

Related model and demo

• —SmolGPT-Fables
• —SmolGPT-Fables Studio

License and provenance

The generated dataset is released under CC BY 4.0. Generator source, quality report, schema, release manifest, and checksums are included for reproducibility. Synthetic output can still reflect assumptions and blind spots encoded by its generator.

Counterfactual copy v5r3

counterfactual-copy-v5r3 is the audited text-only curriculum used by the isolated SmolGPT-Fables SmolLM2 product track. It is additive: the 96,000-row stories config remains the default and is unchanged. The stories config was not mixed into this fine-tune.

• —1,024 rows in 256 split-isolated counterfactual groups of four reskins.
• —816 training rows, 104 validation rows, and 104 test rows.
• —Generator version 5 and curriculum version v5r3.
• —Dataset semantic SHA-256: 6d74c6d1b2c42926adf8a6dfb42f1747fd216489151c1bf5066d689a7a324adb.
• —Annotations SHA-256: 6ab66a657e37ffcbc83a4b58a8244458325d8df9ee38c43a16398e500ad10bb2.
• —Image-generation fields and absolute local paths are excluded.

from datasets import load_dataset

copy_stories = load_dataset(
    "neonforestmist/smolgpt-markdown-stories",
    "counterfactual-copy-v5r3",
)

For reproducible training, pin the immutable Hub commit returned when this config is merged. Config-specific manifest, schema, quality report, and checksums live under metadata/counterfactual-copy-v5r3/. The root dataset_manifest.json, schema.json, and quality-report.json continue to describe the original stories config. The repository's CC BY 4.0 license applies to both configs.