GoldiBench / README.md
VanWang's picture
Add files using upload-large-folder tool
ffd3ed2 verified
|
Raw History Blame Contribute Delete
4.93 kB
metadata
license: cc-by-4.0
pretty_name: GoldiBench
task_categories:
  - image-to-image
tags:
  - agents
  - gui-agents
  - image-editing
  - benchmark
size_categories:
  - n<1K

GoldiBench: Right Tools, Wrong Amount

500 editing tasks in five professional applications (darktable, RawTherapee, GIMP, Inkscape, Blender). Each task pairs a START image with a TARGET rendered from a hidden program, so the amount of every tool has an exact answer. Tasks are posed at three levels: L1 tools and values, L2 tools only, L3 goal only; solvable by writing a program or through the GUI.

Pass rule: closeness = 1 - dE(result, TARGET) / dE(START, TARGET) (mean CIEDE2000, renders <= 1024 px); pass when closeness >= 0.75 and dE <= 4.

Use: hf download VanWang/GoldiBench --repo-type dataset --local-dir data then export GOLDIBENCH_DATA=$PWD/data (see the code repository). Paths inside the files are relative to the dataset root.

Integrity: bench_b1_<app>/FROZEN holds the sha256 of every released program.json; the harness refuses changed tasks. FROZEN.original is the file frozen during the study; the released programs differ from those only by the removed machine-specific prefix of file paths (verified for all 500 tasks).

Files

Same layout as the code expects under GOLDIBENCH_DATA:

bench_b1_<app>/                    program channel, app in darktable rawtherapee gimp inkscape blender
  tasks.jsonl                      one task per line (100): task_id, app, kind (single / multi / long), program (the hidden
                                   answer {tool: {field: value}}), modules (the task's tools), order, intervals (accepted
                                   value ranges), intent (the look in words), l1_instruction (L1 text), source / source_path
                                   (input image or scene), start_target_de00, checks (verification results), generator,
                                   reasons, manual_pages, attempts, revisions
  FROZEN                           sha256 of every task program; runs refuse changed or unlisted tasks
  tasks/<task_id>/start.png        START render (the input as the agent sees it)
  tasks/<task_id>/target.png       TARGET render of the hidden program
  tasks/<task_id>/program.json     the hidden program
  tasks/<task_id>/_val_*.png       validation renders: oracle replay, x0.5 (half), x0.9, x1.1, over-strength
  verify.json                      per-task verification (oracle, strength scaling, single-tool removal)
  magnitude_curve.json             closeness vs. strength scaling of every task (Fig. strength)
  v3_more.json, search.json        extra checks / numeric-search reference (darktable)
  knowledge/                       pre-rendered effect sheets and difference images of the L2 interventions
bench_b1_<app>_gui/                GUI channel
  tasks.jsonl                      one GUI task per line and level (id = <task_id>_<level>): instruction, config (file to
                                   open), allowed_modules, program / gui_program, target, evaluator, kind, level
  tasks/<task_id>/                 start.png / target.png used by the GUI evaluator
  replay_validation*.json          the hidden program replayed through the GUI (task kept only if replay reproduces TARGET)
  rejected.json, unreachable.json  tasks dropped from the GUI channel and why
sources/                           input photos (PSBench) referenced by source_path
assets/                            Blender scenes (Poly Haven) and SVGs (Openclipart) referenced by source_path
manuals/                           application manuals (the --manual intervention); minimanual/: mini manuals and the
                                   task subsets of the Sec. 5 interventions (minimanual/subsets/<app>.txt)

License and sources

CC-BY-4.0 for the tasks, programs, renders and annotations. Input images keep their own licenses: PSBench photos (sources/PSBench-data, see the PSBench license), Poly Haven scenes and HDRIs (CC0, assets/polyhaven), Openclipart drawings (public domain, assets/openclipart). Application manuals (manuals/) belong to their projects (darktable, RawTherapee/RawPedia, GIMP, Inkscape, Blender documentation; see each manual's license).

Citation

@misc{yang2026goldibench,
  title  = {GoldiBench: Right Tools, Wrong Amount -- Benchmarking Agents on Effect Judgement in Professional Editing},
  author = {Yang, Wang and Li, Xinpeng and Zhang, Yanyan and Song, Chaoda and Wang, Shouren and Ganguly, Debargha and
            Singh, Vikash and Ma, Chuang and Cai, Xiaoran and Ye, Kai and Han, Xiaotian and Chaudhary, Vipin},
  year   = {2026}
}