DATASHEET.md

# DECKEDIT-BENCH — Datasheet

Following the *Datasheets for Datasets* structure (Gebru et al., 2021).

## Motivation

Instruction-guided document editing is largely evaluated on plain text or
single-shape image edits. Real presentation editing is harder: instructions are
under-specified, targets must be located among many shapes, edits repeat across
slides, and a good system must change only what was asked while preserving
everything else. DECKEDIT-BENCH provides real decks and natural-language
instructions to measure these abilities, with an explicit axis for **preservation**
(non-target damage) alongside instruction-following accuracy.

## Composition

- **Instances.** 183 `(deck, instruction)` pairs over 28 `.pptx` decks.
- **Deck tiers** (by slide count): Short ≤10 (9 decks), Medium 11–30 (12),
  Long >30 (7). Largest deck: 50 slides.
- **Instructions per tier:** Short 54 · Medium 81 · Long 48.
- **Cells:** ES 17 · EC 57 · PS 87 · PC 22 (see TAXONOMY.md).
- **Languages:** source decks are en (21) / ko (7); some instructions ask for
  translation into en/ko/ja/fr/es/zh.
- **Domains** (28 decks): AI/ML/Speech-NLP research (12), STEM lecture &
  engineering (4), Humanities & general (4), Business reports (3),
  Architecture/Urban/Visual (3), Medical/Bio (2).
- **Visuals:** decks variously contain pictures, tables, and charts; per-deck
  inventory is in `deck_info.json` (`visuals`).
- **Two fully-synthetic decks** (`02_RemoteWorkSurvey`, `09_ApexAnalytics`) are
  chart-heavy fixtures authored for this benchmark.

Each instance carries: `deck_id`, `id`, `cell`, `ops_per_slide`, `actions`,
`derived_from` (template id or `deck-specific`), `intent`, `text`.

### Known imbalances / limitations

- REPLACE accounts for ~75 % of actions; ADD/DELETE/SLIDE are sparser.
- `SLIDE` actions appear only in PS cells; ES/EC/PC × SLIDE = 0.
- `slide-mgmt` (add/delete/duplicate slide) is barely tested.
- Multi-action instructions (≥2 distinct actions) are rare.
- Domain skew toward AI/Speech research (~43 % of decks); legal/finance/sales/
  medical-chart are under-represented — an external-validity caveat.
- Not every instruction is deterministically checkable: ~75 % are
  (text/color/font/chart-attr/background); translate/summarize/"important parts"
  require an LLM/vision judge.

## Collection process

Decks are real lecture, seminar, paper-review, proposal, and report
presentations contributed by their authors (with permission), plus two
synthetic chart decks. Instructions were authored for this benchmark to cover
the Cell × Action × Target space; 92 derive from 14 reusable templates
(`taxonomy.json`), 91 are deck-specific. No model outputs are included in the
released data.

## Preprocessing / cleaning / anonymization

All decks were sanitized for personal and third-party data. PII in `.pptx`
hides in **non-visible metadata layers**; every layer below was scrubbed and the
result re-verified (0 residual names / emails / template watermarks across all
28 decks):

1. **Document properties** — `docProps/core.xml`, `app.xml`: `creator`,
   `lastModifiedBy`, `title`, `subject`, `keywords`, `Company`, `Manager` cleared.
2. **Comment / coauthor records** — `ppt/authors.xml`, `ppt/commentAuthors.xml`:
   real author names anonymized to `Author A/B/…` (5 decks), initials blanked.
3. **Coauthoring revision history** — `ppt/changesInfos/changesInfo*.xml`:
   `chgData name=` author names anonymized (4 decks).
4. **Embedded emails** — real addresses replaced with `example.*` placeholders.

`11_FourthIndustryStartup` was originally built on a commercial third-party
template; it was **rebuilt copyright-clean**: original text and styling preserved
verbatim, the template's full-slide background and all raster images removed,
template-vendor watermarks stripped, and slide imagery replaced with
license-clean Unsplash photos (then downsized to 1920 px). Its instruction
targets (team identifier, author list, four category labels, a replaceable
image, the startup-pitch body) are preserved, so its prompts behave as before.

## Uses

Intended for evaluating instruction-guided `.pptx` editing systems
(instruction-following accuracy, target localization, pattern compression,
and content preservation). Not intended as training data for, or evaluation of, deck
*generation* from scratch. The decks reflect their source domains and languages
and should not be treated as a representative sample of all presentations.

## Distribution

Released as static files (decks + `prompts.jsonl` + `taxonomy.json` +
`deck_info.json` + docs). Decks retain original-resolution media (~1 GB total)
and require Git LFS or a large-file dataset host. License: **CC BY-NC 4.0**
(`LICENSE`); third-party image/figure attributions in `THIRD_PARTY_NOTICES.md`.
This v1 release is **data only** — the evaluation/judge harness is not included.

## Maintenance

When adding decks or instructions: update `deck_info.json`; add the
instructions to the per-deck source and regenerate `prompts.jsonl` /
`taxonomy.json`; re-run the PII scrub and coverage audit. The benchmark is
intended to grow toward the gaps listed under *Known imbalances*.