processbench commited on
Commit
9bf6a16
·
verified ·
1 Parent(s): dd8ecac

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +50 -39
README.md CHANGED
@@ -1,15 +1,16 @@
 
1
  ---
2
  license: other
3
- language:
4
  - en
5
- pretty_name: ProcessBench
6
- task_categories:
7
  - visual-question-answering
8
  tags:
9
  - robotics
10
- - vision-language-models
11
  - embodied-ai
12
  - benchmark
 
13
  - process-understanding
14
  - manipulation
15
  size_categories:
@@ -23,53 +24,63 @@ configs:
23
  path: data/processdata_sft.parquet
24
  ---
25
 
26
- # ProcessBench
27
 
28
  ## Dataset Summary
29
- 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.
30
 
31
- ## What is included
32
- - Frozen evaluation QA items
33
- - SFT QA items or manifests
34
- - Split metadata
35
- - Task distribution statistics
36
- - Human audit summary
37
- - ProcessEval-7B bootstrap CI
38
- - Rendered task cards
39
- - Reconstruction metadata
40
 
41
- ## What is not included
42
- 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.
43
 
44
- ## Data Sources
45
- GM-100, RH20T, REASSEMBLE, and AIST-Bimanual.
 
 
 
 
 
46
 
47
- ## Task Families
48
- T1--T12 with short descriptions.
49
 
50
- ## Data Fields
51
- Explain item_id, source, task_id, input_type, question, choices, answer, visual_ref, source_episode_ref, builder_version.
 
 
 
 
52
 
53
- ## Splits
54
- Strict episode / recording / scene isolation. Report 48,841 SFT and 9,051 eval.
 
 
 
 
55
 
56
- ## Metrics
57
- Accuracy, random baseline, majority baseline, bootstrap CI.
58
 
59
- ## Human Audit
60
- 50 sampled items per task, two annotators.
 
 
 
 
 
 
 
 
 
 
61
 
62
- ## Intended Use
63
- Diagnostic evaluation of VLM-side process understanding and process-aware VLM adaptation.
64
 
65
- ## Out-of-Scope Use
66
- Not a closed-loop robot safety benchmark, not a deployment certificate, not a generic VLM leaderboard.
 
67
 
68
- ## Licenses and Terms
69
- Code MIT; derived data subject to upstream terms.
70
 
71
- ## Croissant Metadata
72
- This release includes Croissant metadata with core and Responsible AI fields.
73
 
74
- ## Citation
75
- Anonymous during review.
 
 
1
+
2
  ---
3
  license: other
4
+ language:
5
  - en
6
+ pretty_name: ProcessBench-v1
7
+ task_categories:
8
  - visual-question-answering
9
  tags:
10
  - robotics
 
11
  - embodied-ai
12
  - benchmark
13
+ - vision-language-models
14
  - process-understanding
15
  - manipulation
16
  size_categories:
 
24
  path: data/processdata_sft.parquet
25
  ---
26
 
27
+ # ProcessBench-v1
28
 
29
  ## Dataset Summary
 
30
 
31
+ ProcessBench is a process-aware benchmark for robotic manipulation understanding. This release build contains `57,892` public QA rows: `9,051` eval rows and `48,841` SFT rows across `12` task families.
 
 
 
 
 
 
 
 
32
 
33
+ ## Included in this local release build
 
34
 
35
+ - `data/processbench_eval.parquet` and `data/processbench_eval.jsonl`
36
+ - `data/processdata_sft.parquet` and `data/processdata_sft.jsonl`
37
+ - dataset-specific split files for all four sources
38
+ - `metadata/split_summary.json`, `metadata/eval_manifest.json`, `metadata/sft_manifest.json`
39
+ - `metadata/task_distribution.csv`
40
+ - `metadata/schema.md`, `metadata/reconstruction.md`, `metadata/prompt_templates.md`
41
+ - post-trained `ProcessEval-7B` results under `SFT_results/`
42
 
43
+ ## Not included in this local release build
 
44
 
45
+ - full upstream raw videos
46
+ - full extracted frame caches
47
+ - local absolute paths from the development machine
48
+ - separate structured task-meta fields
49
+ - Croissant metadata
50
+ - task-card PNGs and human-audit assets
51
 
52
+ ## Sources
53
+
54
+ - `GM-100`
55
+ - `RH20T`
56
+ - `REASSEMBLE`
57
+ - `AIST-Bimanual`
58
 
59
+ ## Task Families
 
60
 
61
+ - `T1`: Phase Recognition
62
+ - `T2`: Contact Detection
63
+ - `T3`: Motion Direction Prediction
64
+ - `T4`: Bimanual Coordination State
65
+ - `T5`: Primitive-local Progress
66
+ - `T6`: Motion State Recognition
67
+ - `T7`: Operation Outcome Prediction
68
+ - `T8`: Temporal Ordering
69
+ - `T9`: Temporal Priority Prediction
70
+ - `T10`: Current Primitive Recognition
71
+ - `T11`: Next Primitive Prediction
72
+ - `T12`: Primitive Chain Restoration
73
 
74
+ ## Splits
 
75
 
76
+ - `eval`: `9,051`
77
+ - `sft`: `48,841`
78
+ - split rule: strict episode / recording / scene isolation
79
 
80
+ ## License and Terms
 
81
 
82
+ This release uses `license: other`.
 
83
 
84
+ - Derived benchmark metadata in this release remains subject to upstream dataset terms.
85
+ - Raw videos and full frame caches are not redistributed here.
86
+ - Full visual reconstruction requires obtaining the upstream datasets under their original terms.