Datasets:
license: other
language:
- en
pretty_name: ProcessBench
task_categories:
- visual-question-answering
tags:
- robotics
- vision-language-models
- embodied-ai
- benchmark
- process-understanding
- manipulation
size_categories:
- 10K<n<100K
configs:
- config_name: default
data_files:
- split: eval
path: data/processbench_eval.parquet
- split: train
path: data/processdata_sft.parquet
ProcessBench
Dataset Summary
ProcessBench is a process-aware VLM benchmark for robotic manipulation. It contains 57,892 QA items, including 48,841 SFT items and 9,051 evaluation items, across 12 task families.
What is included
- Frozen evaluation QA items
- SFT QA items or manifests
- Split metadata
- Task distribution statistics
- Human audit summary
- ProcessEval-7B bootstrap CI
- Rendered task cards
- Reconstruction metadata
What is not included
We do not redistribute full upstream raw videos or full frame dumps. Full visual reconstruction requires users to obtain the upstream datasets under their original licenses.
Data Sources
GM-100, RH20T, REASSEMBLE, and AIST-Bimanual.
Task Families
T1--T12 with short descriptions.
Data Fields
Explain item_id, source, task_id, input_type, question, choices, answer, visual_ref, source_episode_ref, builder_version.
Splits
Strict episode / recording / scene isolation. Report 48,841 SFT and 9,051 eval.
Metrics
Accuracy, random baseline, majority baseline, bootstrap CI.
Human Audit
50 sampled items per task, two annotators.
Intended Use
Diagnostic evaluation of VLM-side process understanding and process-aware VLM adaptation.
Out-of-Scope Use
Not a closed-loop robot safety benchmark, not a deployment certificate, not a generic VLM leaderboard.
Licenses and Terms
Code MIT; derived data subject to upstream terms.
Croissant Metadata
This release includes Croissant metadata with core and Responsible AI fields.
Citation
Anonymous during review.