benchmark_card.md

ProcessBench Benchmark Card

Positioning

ProcessBench is a signal-grounded, process-aware benchmark for robotic manipulation understanding.

Current public release contents

• merged eval and SFT tables in JSONL and Parquet
• dataset-specific eval and SFT split files for all four sources
• split manifests and source-level split counts
• task-distribution statistics
• SFT results for qwen2.5-vl-7b-instruct
• schema, reconstruction, and prompt-template docs
• full ProcessData-58k suite
• LoRA adapter weights and configuration for ProcessEval-7B

Source roles

Source	Role	Supported tasks	Frozen size
GM-100	Main goal-conditioned source	T1, T2, T3, T4, T5, T6, T8, T9	15,500
RH20T	Force/torque and contact source	T1, T2, T3, T5, T6, T7, T8, T9	15,800
REASSEMBLE	Action-chain and primitive source	T1, T2, T5, T6, T7, T8, T9, T10, T11, T12	17,165
AIST-Bimanual	Bimanual kinematics source	T3, T4, T6, T8, T9	9,427

Split protocol

• eval ratio: 15%
• isolation unit: episode or recording depending on source
• totals: 48,841 SFT and 9,051 eval