README.md

---
language:
- en
license: cc0-1.0
pretty_name: Screenplay Format Edge Cases
size_categories:
- n<1K
task_categories:
- text-classification
tags:
- screenwriting
- evaluation
- synthetic
- inkwell
configs:
- config_name: default
  data_files:
  - split: test
    path: data/cases.jsonl
---
# Screenplay Format Edge Cases

48 original Fountain specimens in 24 contrast pairs — two near-identical inputs
per pair, at the points where the Fountain syntax leaves a choice. In 17 pairs
the one difference changes how the lines are classified. In the other 7 it
changes the surface and the labels hold: a lowercase scene prefix, a cue
extension, a non-Latin cue, escaped characters, a dual-dialogue caret, an inline
note and centered-text markers.

**Version:** 1.0.0 · **Maintainer:** Inkwell.

Built by [Inkwell](https://www.inkwell.software/?utm_source=huggingface&utm_medium=referral&utm_campaign=observatory&utm_content=screenplay_format_edge_cases_card_home), the IDE for screenwriters.
Paste Fountain into the [Fountain validator](https://www.inkwell.software/tools/fountain-validator?utm_source=huggingface&utm_medium=referral&utm_campaign=observatory&utm_content=screenplay_format_edge_cases_card_validator)
to see the scenes, character cues, dialogue and action it recognizes.

## What's included

`data/cases.jsonl` holds all 48 records; `schema.json` is a JSON Schema for one
record. Each record carries the Fountain input, its expected line-level element
types, the contrast pair it belongs to, and a short rationale for the expected
labels.

## Provenance

Every specimen is original Fountain written for Inkwell and released under
CC0-1.0.

## How the labels work

The expected labels are line-level screenplay element types. A pair shares a
`pair_id`, and `split_group` repeats it: keep both variants of a pair in the same
split, because the point of a pair is the single difference between its two
inputs.

Forty-two of the 48 specimens fall inside the subset a demonstration Fountain
parser implements, and `supported_by_demo` marks which. The other six use dual
dialogue, a note, boneyard, a section, lyrics and centered text — valid Fountain
that the demonstration exporter stops on rather than guessing. The flag records
one implementation's coverage.

Two more things the labels assume: an FDX paragraph can group adjacent action or
dialogue lines, so a paragraph count can differ from the line-level label count;
and a forced-element marker (`!`, `@`, `.`, `>`) is syntax, not rendered text.

## What it's for

Teaching the syntax contrasts, building regression tests, and watching which
single edit changes how a line is classified. Extend it with your own cases as
your parser grows.

## Load the records

```python
from datasets import load_dataset
cases = load_dataset("Inkwell-Software/screenplay-format-edge-cases", split="test")
print(cases[0])
```

## Citation

Cite "Inkwell, Screenplay Format Edge Cases, version 1.0.0"; the machine-readable
form is in [`CITATION.cff`](./CITATION.cff).

## Maintenance

Open a discussion with the case ID, the expected behavior, the observed behavior,
and the relevant part of the [Fountain spec](https://fountain.io/syntax/). A
change to expected labels is a version increment and a changelog entry.