| --- |
| license: cc0-1.0 |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - narrative |
| - long-form-generation |
| - knowledge-graph |
| - story-generation |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # DSG State-Continuation |
|
|
| Training data for **graph-conditioned long-form fiction generation**: given the |
| narrative state a reader would hold after chapters 1..t-1, and a one-line brief |
| for chapter t, write chapter t. |
|
|
| Built from **215 public-domain novels** (Project Gutenberg, English fiction), |
| segmented into chapters. **6,876 examples.** |
|
|
| ## Why the state is built this way |
|
|
| The state is not a summary and not a retrieval index. It is a revision-aware |
| assertion store built **causally** — at example *t* it contains only what a |
| reader could know from chapters 1..t-1 — and maintained by an explicit update |
| calculus that distinguishes three things append-only pipelines conflate: |
|
|
| | operation | when | effect | |
| |---|---|---| |
| | `ELABORATE` | later text specifies what earlier text left open | refined in place, nothing retracted | |
| | `SUPERSEDE` | the **story world** changed | validity interval closed, kept as history | |
| | `REVISE` | the **reader** was wrong; it was never true | retracted, and so is anything derived from it | |
|
|
| The `SUPERSEDE`/`REVISE` split is the operational form of the narratological |
| distinction between an event in the story and a disclosure in the telling. It is |
| decided by an inspectable lexicon — predicate mutability, evidential source, |
| revelation markers — never by a language model. |
|
|
| On 28 annotated novels this calculus holds **0.0%** of the state's slots |
| self-contradictory, against **37–42%** for an append-only store, at every model |
| size tested (1.5B / 3B / 7B). |
|
|
| ## Fields |
|
|
| | field | description | |
| |---|---| |
| | `book_id`, `title`, `author` | Gutenberg source | |
| | `chapter` | 1-indexed chapter number | |
| | `state` | the graph state after chapters 1..t-1, as a bounded fact digest | |
| | `beat` | one-sentence brief for chapter t | |
| | `target` | the real chapter text (mean ~4,100 characters) | |
| | `target_chars` | length of the target | |
|
|
| 3.2% of rows have an empty state — these are chapter 1, the cold start. |
|
|
| ## Intended use |
|
|
| Fine-tuning a writer to *use* a maintained narrative state rather than merely |
| have one in context. The point is scaling: a state digest does not grow with the |
| story, while a transcript does. |
|
|
| ## Limitations |
|
|
| - **English public-domain fiction**, mostly 19th–early 20th century. Prose style |
| and naming conventions are period-specific. |
| - States are **extracted, not annotated**: produced by Qwen2.5-7B-Instruct and |
| reconciled symbolically. They contain extraction noise, including invented |
| predicates and occasional wrong attributions. |
| - Chapter segmentation is heuristic (headings, with a paragraph-block fallback). |
| - Splitting for training should be **by book**, not by row: chapters from one |
| novel share state and leak across a random split. |
|
|
| ## Source |
|
|
| Code, calculus and evaluation: the Dynamic Story Graph project. |
| Corpus: Project Gutenberg via `sedthh/gutenberg_english`. |
|
|
