Datasets:

EditPPT
/

DECKEDIT-BENCH

pretty_name: DECKEDIT-BENCH
license: cc-by-nc-4.0
language:
  - en
  - ko
tags:
  - powerpoint
  - pptx
  - instruction-following
  - document-editing
  - agents
  - benchmark
size_categories:
  - n<1K
task_categories:
  - other
configs:
  - config_name: prompts
    data_files: prompts.jsonl

DECKEDIT-BENCH

A benchmark for instruction-guided PowerPoint (.pptx) editing. It pairs 28 real-world presentation decks with 183 natural-language editing instructions, organized along a Cell × Action × Target taxonomy and three deck-length tiers (Short / Medium / Long).

Each instance is a (deck, instruction) pair: a system reads the instruction, edits the deck, and the before/after decks are compared. The benchmark stresses locating the right targets, decomposing multi-step instructions, compressing repeated edits into a pattern, and not damaging what should stay untouched.

DECKEDIT-BENCH/
├── decks/{Short,Medium,Long}/*.pptx   # 28 source decks (≤10 / 11–30 / >30 slides)
├── prompts.jsonl                       # 183 editing instructions (one JSON / line)
├── taxonomy.json                       # cell + action definitions, 14 reusable templates
├── deck_info.json                      # per-deck metadata (size, language, topic, visuals)
├── DATASHEET.md                        # datasheet-for-datasets (composition, collection, cleaning)
├── TAXONOMY.md                         # Cell × Action × Target taxonomy + template catalog
├── THIRD_PARTY_NOTICES.md              # image/figure provenance & attributions
├── LICENSE                             # data license (CC BY-NC 4.0)
└── CITATION.cff

Prompt format (`prompts.jsonl`)

One JSON object per line:

{"deck_id":"02_UTI","id":"02_UTI-01","cell":"PS","ops_per_slide":1,
 "actions":["REPLACE"],"derived_from":"U-01",
 "intent":"…what this instance tests…","text":"…instruction sent to the system…"}

text — the instruction handed to the editing system.
cell ∈ {ES, EC, PS, PC} · actions ⊆ {ADD, DELETE, REPLACE, SLIDE}.
derived_from — a universal template id (U-xx, see taxonomy.json) or "deck-specific".

See TAXONOMY.md for definitions.

Quickstart

import json
prompts = [json.loads(l) for l in open("prompts.jsonl", encoding="utf-8")]
info = {d["id"]: d for d in json.load(open("deck_info.json", encoding="utf-8"))["decks"]}

p = prompts[0]
deck_path = info[p["deck_id"]]["path"]      # e.g. "decks/Short/01_CausalAGI.pptx"
# 1. open deck_path, 2. apply p["text"] with your system, 3. diff before/after.

Coverage

	total	breakdown
Decks	28	Short 9 · Medium 12 · Long 7
Prompts	183	Short 54 · Medium 81 · Long 48
Cells	—	ES 17 · EC 57 · PS 87 · PC 22
Deck languages	—	en 21 · ko 7 (translate targets also span ja/fr/es/zh)

Full Cell × Action matrix and known imbalances are in DATASHEET.md.

Notes on the decks

Full resolution. The .pptx files keep original-resolution media (~1 GB total) and are distributed via Git LFS / the dataset host.
Privacy. Every deck was scrubbed of author/coauthoring/comment PII; see DATASHEET.md → Preprocessing/cleaning.
Provenance. Embedded figures in paper-review decks, and the replacement images in 11_FourthIndustryStartup, are credited in THIRD_PARTY_NOTICES.md.

License & citation

Data is released under CC BY-NC 4.0 (see LICENSE). If you use DECKEDIT-BENCH, please cite it — see CITATION.cff.