processbench commited on
Commit
1f69c5b
·
verified ·
1 Parent(s): e25596c

Upload 6 files

Browse files
metadata/eval_manifest.json ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "splits_v1",
3
+ "builder_version": "processbench_v1",
4
+ "visual_inputs_redistributed": false,
5
+ "reconstruction_required": true,
6
+ "merged_eval_jsonl": "data/processbench_eval.jsonl",
7
+ "merged_eval_parquet": "data/processbench_eval.parquet",
8
+ "datasets": [
9
+ {
10
+ "source": "GM-100",
11
+ "slug": "gm100",
12
+ "eval_jsonl": "data/gm100_eval.jsonl",
13
+ "num_items_eval": 2643,
14
+ "num_groups_eval": 628,
15
+ "random_baselines": {
16
+ "T1": 0.25,
17
+ "T2": 0.5,
18
+ "T3": 0.25,
19
+ "T4": 0.25,
20
+ "T6": 0.5,
21
+ "T9": 0.5,
22
+ "T5": 0.3333333333333333,
23
+ "T8": 0.16666666666666666
24
+ }
25
+ },
26
+ {
27
+ "source": "RH20T",
28
+ "slug": "rh20t",
29
+ "eval_jsonl": "data/rh20t_eval.jsonl",
30
+ "num_items_eval": 2422,
31
+ "num_groups_eval": 89,
32
+ "random_baselines": {
33
+ "T1": 0.25,
34
+ "T2": 0.5,
35
+ "T3": 0.25,
36
+ "T6": 0.5,
37
+ "T7": 0.5,
38
+ "T9": 0.5,
39
+ "T5": 0.3333333333333333,
40
+ "T8": 0.16666666666666666
41
+ }
42
+ },
43
+ {
44
+ "source": "REASSEMBLE",
45
+ "slug": "reassemble",
46
+ "eval_jsonl": "data/reassemble_eval.jsonl",
47
+ "num_items_eval": 2562,
48
+ "num_groups_eval": 6,
49
+ "random_baselines": {
50
+ "T1": 0.25,
51
+ "T10": 0.25,
52
+ "T11": 0.25,
53
+ "T12": 0.25,
54
+ "T2": 0.5,
55
+ "T6": 0.5,
56
+ "T7": 0.5,
57
+ "T9": 0.5,
58
+ "T5": 0.3333333333333333,
59
+ "T8": 0.16666666666666666
60
+ }
61
+ },
62
+ {
63
+ "source": "AIST-Bimanual",
64
+ "slug": "aist",
65
+ "eval_jsonl": "data/aist_eval.jsonl",
66
+ "num_items_eval": 1424,
67
+ "num_groups_eval": 30,
68
+ "random_baselines": {
69
+ "T3": 0.25,
70
+ "T4": 0.25,
71
+ "T6": 0.5,
72
+ "T9": 0.5,
73
+ "T8": 0.16666666666666666
74
+ }
75
+ }
76
+ ]
77
+ }
metadata/prompt_templates.md ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ # Prompt Templates
3
+
4
+ This file documents the public prompt serialization used by the current release build. Separate source-side task-meta prompting is intentionally disabled in this public release.
5
+
6
+ ## Standard multiple-choice template
7
+
8
+ ```text
9
+ {question}
10
+
11
+ A: {choice_A}
12
+ B: {choice_B}
13
+ ...
14
+
15
+ Choose exactly one option label.
16
+
17
+ Output protocol:
18
+ - You may include brief reasoning before the final answer.
19
+ - The final line must be exactly: <ANSWER>A</ANSWER>
20
+ - Do not output anything after </ANSWER>.
21
+ ```
22
+
23
+ ## `T8` Temporal Ordering template
24
+
25
+ ```text
26
+ You are shown 3 frames from a robot manipulation task.
27
+ The frames are labeled X, Y, Z (these labels are arbitrary identifiers, not positional, and not time-ordered).
28
+ Determine the correct chronological order of these frames (from earliest to latest).
29
+ Choose exactly one 3-letter permutation, for example: YXZ means Y happened first, then X, then Z.
30
+
31
+ Output protocol:
32
+ - You may include brief reasoning before the final answer.
33
+ - The final line must be exactly: <ANSWER>XYZ</ANSWER>
34
+ - Do not output anything after </ANSWER>.
35
+ ```
36
+
37
+ ## `T9` Temporal Priority template
38
+
39
+ ```text
40
+ A single comparison image shows two labeled robot-manipulation panels from the same episode.
41
+ The left-right placement of the panels and the labels are arbitrary identifiers and do not indicate temporal order.
42
+ Which labeled panel happened earlier in the real manipulation sequence, X or Y?
43
+ Choose exactly one label: X or Y.
44
+
45
+ Output protocol:
46
+ - You may include brief reasoning before the final answer.
47
+ - The final line must be exactly: <ANSWER>X</ANSWER>
48
+ - Do not output anything after </ANSWER>.
49
+ ```
50
+
51
+ ## Public release note
52
+
53
+ The public release tables normalize all choices to flat columns and do not expose separate structured task-meta fields.
metadata/reconstruction.md ADDED
@@ -0,0 +1,42 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ # Reconstruction
3
+
4
+ ## What this release contains
5
+
6
+ - Public QA tables for `eval` and `sft`.
7
+ - Split manifests and split counts.
8
+ - Public release-relative references for each item.
9
+
10
+ ## What this release does not contain
11
+
12
+ - Full raw videos.
13
+ - Full extracted frame caches.
14
+ - Local absolute paths used during internal development.
15
+ - Separate structured task-meta fields.
16
+
17
+ ## How to use the public references
18
+
19
+ Each row exposes three reconstruction-oriented fields:
20
+
21
+ - `visual_ref`
22
+ - `source_episode_ref`
23
+ - `reconstruction_key_json`
24
+
25
+ These are designed to be stable, path-free references. They are sufficient to map a released item back to the corresponding source episode or recording once you have obtained the upstream dataset under its original terms.
26
+
27
+ ## Source-specific units
28
+
29
+ - `GM-100`: grouped by `(task_id, episode_id)`.
30
+ - `RH20T`: grouped by `recording_id`.
31
+ - `REASSEMBLE`: grouped by `recording_id`.
32
+ - `AIST-Bimanual`: grouped by `recording_id`.
33
+
34
+ ## Expected workflow
35
+
36
+ 1. Obtain upstream datasets under their original terms.
37
+ 2. Use `source`, `source_task_id`, `source_unit_id`, `camera`, and `frame_indices_json` from the public release.
38
+ 3. Reconstruct the required visual inputs with the source-specific evaluation scripts in the main codebase.
39
+
40
+ ## Release policy
41
+
42
+ This Hugging Face release is a reviewable minimal closed loop for metadata, QA, splits, statistics, and baseline results. Full visual reconstruction remains upstream-data dependent.
metadata/schema.md ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ # Schema
3
+
4
+ ## Public QA fields
5
+
6
+ - `item_id`: Stable public item identifier.
7
+ - `split`: `eval` or `sft`.
8
+ - `source`: `GM-100`, `RH20T`, `REASSEMBLE`, or `AIST-Bimanual`.
9
+ - `source_slug`: Short source slug used in release-relative references.
10
+ - `source_task_id`: Source-side task identifier without local paths.
11
+ - `source_unit_type`: Split-isolation unit type, e.g. `episode` or `recording`.
12
+ - `source_unit_id`: Public unit identifier without local paths.
13
+ - `task_id`: Canonical paper task ID `T1` ... `T12`.
14
+ - `task_name`: Canonical paper task name.
15
+ - `task_type_legacy`: Legacy engineering task name when it differs from paper naming.
16
+ - `input_type`: Public input format description.
17
+ - `question`: Public question text.
18
+ - `choice_A` ... `choice_F`: Flattened answer choices.
19
+ - `answer`: Public answer label in the release schema.
20
+ - `answer_text`: Public answer text mapped from `answer`.
21
+ - `num_choices`: Number of non-null answer choices.
22
+ - `num_frames`: Number of referenced frames.
23
+ - `frame_indices_json`: JSON-encoded frame indices.
24
+ - `display_labels_json`: JSON-encoded panel labels such as `["X", "Y"]` or `["X", "Z", "Y"]`.
25
+ - `camera`: Source camera identifier when available.
26
+ - `arm_type`: Source-side arm configuration tag when available.
27
+ - `visual_ref`: Release-relative visual reference string. This is not a local filesystem path.
28
+ - `source_episode_ref`: Release-relative source unit reference.
29
+ - `reconstruction_key_json`: Minimal JSON payload needed to reconstruct the sample with upstream data and source-specific scripts.
30
+ - `task_meta_in_source`: Whether source-side task metadata/context exists upstream.
31
+ - `task_meta_public`: Always `false` in this release build. Separate structured task-meta fields are intentionally withheld.
32
+ - `split_version`: Current split version.
33
+ - `split_group_id`: Source-side split group identifier.
34
+ - `builder_version`: Public builder version tag.
35
+ - `prompt_version`: Public prompt-serialization version tag.
36
+ - `sft_target`: SFT-only supervised target string. Present only in `processdata_sft.*`.
37
+
38
+ ## Notes
39
+
40
+ - `T5` is equivalent to `T_progress`.
41
+ - `T8` is equivalent to `T_temporal`.
42
+ - `T9` is equivalent to `T_binary`.
43
+ - `T8` is normalized into six explicit permutation choices so the public table remains flat and Parquet-friendly.
44
+ - `T9` is normalized to `A/B` choices even when the source data originally used `X/Y`.
metadata/sft_manifest.json ADDED
@@ -0,0 +1,38 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "splits_v1",
3
+ "builder_version": "processbench_v1",
4
+ "visual_inputs_redistributed": false,
5
+ "reconstruction_required": true,
6
+ "merged_sft_jsonl": "data/processdata_sft.jsonl",
7
+ "merged_sft_parquet": "data/processdata_sft.parquet",
8
+ "datasets": [
9
+ {
10
+ "source": "GM-100",
11
+ "slug": "gm100",
12
+ "sft_jsonl": "data/gm100_sft.jsonl",
13
+ "num_items_sft": 12857,
14
+ "num_groups_sft": 3558
15
+ },
16
+ {
17
+ "source": "RH20T",
18
+ "slug": "rh20t",
19
+ "sft_jsonl": "data/rh20t_sft.jsonl",
20
+ "num_items_sft": 13378,
21
+ "num_groups_sft": 507
22
+ },
23
+ {
24
+ "source": "REASSEMBLE",
25
+ "slug": "reassemble",
26
+ "sft_jsonl": "data/reassemble_sft.jsonl",
27
+ "num_items_sft": 14603,
28
+ "num_groups_sft": 31
29
+ },
30
+ {
31
+ "source": "AIST-Bimanual",
32
+ "slug": "aist",
33
+ "sft_jsonl": "data/aist_sft.jsonl",
34
+ "num_items_sft": 8003,
35
+ "num_groups_sft": 170
36
+ }
37
+ ]
38
+ }
metadata/split_summary.json ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "splits_v1",
3
+ "builder_version": "processbench_v1",
4
+ "seed": 42,
5
+ "eval_ratio": 0.15,
6
+ "split_policy": "strict episode / recording / scene isolation",
7
+ "task_aliases": {
8
+ "T5": "T_progress",
9
+ "T8": "T_temporal",
10
+ "T9": [
11
+ "T_binary",
12
+ "T9"
13
+ ]
14
+ },
15
+ "sft_items": 48841,
16
+ "eval_items": 9051,
17
+ "sources": {
18
+ "GM-100": {
19
+ "slug": "gm100",
20
+ "group_unit": "(task_id, episode_id)",
21
+ "total": 15500,
22
+ "sft": 12857,
23
+ "eval": 2643,
24
+ "eval_groups": 628,
25
+ "sft_groups": 3558,
26
+ "task_counts_total": {
27
+ "T1": 3000,
28
+ "T2": 1000,
29
+ "T3": 2500,
30
+ "T4": 1500,
31
+ "T5": 2500,
32
+ "T6": 1500,
33
+ "T8": 2000,
34
+ "T9": 1500
35
+ },
36
+ "task_counts_eval": {
37
+ "T1": 597,
38
+ "T2": 150,
39
+ "T3": 483,
40
+ "T4": 225,
41
+ "T5": 424,
42
+ "T6": 223,
43
+ "T8": 318,
44
+ "T9": 223
45
+ },
46
+ "task_counts_sft": {
47
+ "T1": 2403,
48
+ "T2": 850,
49
+ "T3": 2017,
50
+ "T4": 1275,
51
+ "T5": 2076,
52
+ "T6": 1277,
53
+ "T8": 1682,
54
+ "T9": 1277
55
+ }
56
+ },
57
+ "RH20T": {
58
+ "slug": "rh20t",
59
+ "group_unit": "recording_id",
60
+ "total": 15800,
61
+ "sft": 13378,
62
+ "eval": 2422,
63
+ "eval_groups": 89,
64
+ "sft_groups": 507,
65
+ "task_counts_total": {
66
+ "T1": 2000,
67
+ "T2": 1500,
68
+ "T3": 1800,
69
+ "T5": 2300,
70
+ "T6": 1500,
71
+ "T7": 4000,
72
+ "T8": 1200,
73
+ "T9": 1500
74
+ },
75
+ "task_counts_eval": {
76
+ "T1": 303,
77
+ "T2": 223,
78
+ "T3": 268,
79
+ "T5": 366,
80
+ "T6": 227,
81
+ "T7": 632,
82
+ "T8": 181,
83
+ "T9": 222
84
+ },
85
+ "task_counts_sft": {
86
+ "T1": 1697,
87
+ "T2": 1277,
88
+ "T3": 1532,
89
+ "T5": 1934,
90
+ "T6": 1273,
91
+ "T7": 3368,
92
+ "T8": 1019,
93
+ "T9": 1278
94
+ }
95
+ },
96
+ "REASSEMBLE": {
97
+ "slug": "reassemble",
98
+ "group_unit": "recording_id",
99
+ "total": 17165,
100
+ "sft": 14603,
101
+ "eval": 2562,
102
+ "eval_groups": 6,
103
+ "sft_groups": 31,
104
+ "task_counts_total": {
105
+ "T1": 2523,
106
+ "T2": 2523,
107
+ "T5": 2931,
108
+ "T6": 1260,
109
+ "T7": 1110,
110
+ "T8": 907,
111
+ "T9": 1849,
112
+ "T10": 2438,
113
+ "T11": 1338,
114
+ "T12": 286
115
+ },
116
+ "task_counts_eval": {
117
+ "T1": 374,
118
+ "T2": 374,
119
+ "T5": 432,
120
+ "T6": 206,
121
+ "T7": 161,
122
+ "T8": 135,
123
+ "T9": 275,
124
+ "T10": 359,
125
+ "T11": 200,
126
+ "T12": 46
127
+ },
128
+ "task_counts_sft": {
129
+ "T1": 2149,
130
+ "T2": 2149,
131
+ "T5": 2499,
132
+ "T6": 1054,
133
+ "T7": 949,
134
+ "T8": 772,
135
+ "T9": 1574,
136
+ "T10": 2079,
137
+ "T11": 1138,
138
+ "T12": 240
139
+ }
140
+ },
141
+ "AIST-Bimanual": {
142
+ "slug": "aist",
143
+ "group_unit": "recording_id",
144
+ "total": 9427,
145
+ "sft": 8003,
146
+ "eval": 1424,
147
+ "eval_groups": 30,
148
+ "sft_groups": 170,
149
+ "task_counts_total": {
150
+ "T3": 3041,
151
+ "T4": 1989,
152
+ "T6": 1597,
153
+ "T8": 2200,
154
+ "T9": 600
155
+ },
156
+ "task_counts_eval": {
157
+ "T3": 456,
158
+ "T4": 308,
159
+ "T6": 240,
160
+ "T8": 330,
161
+ "T9": 90
162
+ },
163
+ "task_counts_sft": {
164
+ "T3": 2585,
165
+ "T4": 1681,
166
+ "T6": 1357,
167
+ "T8": 1870,
168
+ "T9": 510
169
+ }
170
+ }
171
+ },
172
+ "release_files": {
173
+ "merged_eval_jsonl": "data/processbench_eval.jsonl",
174
+ "merged_eval_parquet": "data/processbench_eval.parquet",
175
+ "merged_sft_jsonl": "data/processdata_sft.jsonl",
176
+ "merged_sft_parquet": "data/processdata_sft.parquet"
177
+ }
178
+ }