mcobzarenco commited on
Commit
bf02e46
·
verified ·
1 Parent(s): e37d968

Upload analysis__disc1000_panel_row_audit.json with huggingface_hub

Browse files
analysis__disc1000_panel_row_audit.json ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "npz": "reports/eval__grasp_sft_v2_demosonly_1gpu_disc__step_001000__panel_v2_k4l2_euler10_draws1_stable.npz",
3
+ "checkpoint": "/home/ubuntu/checkpoints/finetune/grasp_sft_v2_demosonly_1gpu_disc/step_001000",
4
+ "wear_facts": {
5
+ "leg_molmo_norm": "checkpoint",
6
+ "flow_normalization_tag": "q01q99",
7
+ "train_per_dataset_flow_norm": false,
8
+ "per_dataset_stats_keys": [
9
+ "grasp_demos_v2/merged"
10
+ ],
11
+ "stats_note": "global stats RECOMPUTED at launch: exact pooled quantiles over 1750071 train-split frames of ['grasp_demos_v2/merged'], oriented like the source table; the original table is the source checkpoint's (/home/ubuntu/checkpoints/molmoact2-so101-released)",
12
+ "worn_table": {
13
+ "mean": [
14
+ 0.09233876317739487,
15
+ -7.418671607971191,
16
+ 22.1364803314209,
17
+ 51.24595642089844,
18
+ 82.55037689208984,
19
+ 12.741997718811035
20
+ ],
21
+ "std": [
22
+ 27.988689422607422,
23
+ 40.790802001953125,
24
+ 31.966609954833984,
25
+ 20.678668975830078,
26
+ 77.26783752441406,
27
+ 19.013696670532227
28
+ ],
29
+ "q01": [
30
+ -54.11045837402344,
31
+ 55.71287155151367,
32
+ -65.48938751220703,
33
+ 11.65783405303955,
34
+ -157.21087646484375,
35
+ 0.0
36
+ ],
37
+ "q99": [
38
+ 50.82916259765625,
39
+ -113.50212097167969,
40
+ 97.0,
41
+ 95.00440979003906,
42
+ 157.21087646484375,
43
+ 41.692020416259766
44
+ ]
45
+ },
46
+ "verdict": "merged scheme: flow_denormalize_chunk(per_dataset=False) \u2014 every item wore the recomputed-at-launch demos-only global table; per_dataset_stats is never consulted at eval"
47
+ },
48
+ "caveats": [
49
+ "re-wear counterfactuals re-express the OUTPUT side only: the state input was normalized/binned through the worn demos table when the model ran, so state-side OOD is not undone \u2014 they bound the wear share, they do not re-run the model",
50
+ "per-repo rows are fit on the panel's own truth (np.quantile linear over valid chunk elements, duplicate-weighted by window overlap) \u2014 oracle-ish rows, optimistic vs a deployment row estimate"
51
+ ],
52
+ "anchors_reproduced": {
53
+ "bijou@1000": 58.143384213383385,
54
+ "state-copy": 8.367757873526733
55
+ },
56
+ "box_audit": {
57
+ "truth_outside_frac_per_joint": {
58
+ "shoulder_pan": 0.07369556616706689,
59
+ "shoulder_lift": 0.7296525803453955,
60
+ "elbow_flex": 0.5511537718402385,
61
+ "wrist_flex": 0.19014923587917598,
62
+ "wrist_roll": 0.027241068389043766,
63
+ "gripper": 0.26894430796691415
64
+ },
65
+ "truth_any_joint_outside_frac": 0.8583248226558834,
66
+ "floor_chunk_mae": 14.401469230651855,
67
+ "floor_per_joint": {
68
+ "shoulder_pan": 1.4592567682266235,
69
+ "shoulder_lift": 52.04825210571289,
70
+ "elbow_flex": 24.017120361328125,
71
+ "wrist_flex": 6.615574359893799,
72
+ "wrist_roll": 0.5258612036705017,
73
+ "gripper": 1.74805748462677
74
+ },
75
+ "pred_edge_saturation_per_joint": {
76
+ "shoulder_pan": 0.000588846872211028,
77
+ "shoulder_lift": 0.0014905186452841646,
78
+ "elbow_flex": 0.00014491153495818266,
79
+ "wrist_flex": 0.002872928685123336,
80
+ "wrist_roll": 0.0016561318280935162,
81
+ "gripper": 0.11585562210751976
82
+ }
83
+ },
84
+ "actual_per_joint_mae": {
85
+ "shoulder_pan": 23.462173461914062,
86
+ "shoulder_lift": 104.43495178222656,
87
+ "elbow_flex": 99.3005599975586,
88
+ "wrist_flex": 27.106822967529297,
89
+ "wrist_roll": 71.17012023925781,
90
+ "gripper": 23.381319046020508
91
+ },
92
+ "rewear": {
93
+ "round_trip_worst_abs_deg": 1.52587890625e-05,
94
+ "repo_rows_chunk_mae": 27.395334001919885,
95
+ "repo_rows_per_joint": {
96
+ "shoulder_pan": 21.10439109802246,
97
+ "shoulder_lift": 41.8028678894043,
98
+ "elbow_flex": 36.7511100769043,
99
+ "wrist_flex": 22.337148666381836,
100
+ "wrist_roll": 22.218639373779297,
101
+ "gripper": 20.153600692749023
102
+ },
103
+ "released_table_chunk_mae": 54.399765688015435,
104
+ "released_table_per_joint": {
105
+ "shoulder_pan": 24.25238609313965,
106
+ "shoulder_lift": 105.2990951538086,
107
+ "elbow_flex": 98.5770263671875,
108
+ "wrist_flex": 27.263837814331055,
109
+ "wrist_roll": 46.14845275878906,
110
+ "gripper": 24.851577758789062
111
+ },
112
+ "null_repo_midpoint_chunk_mae": 25.154475696265422,
113
+ "n_repo_rows": 838
114
+ },
115
+ "demos_prior_collapse": {
116
+ "pred_vs_demos_mean_mae": 22.636736684240066,
117
+ "truth_vs_demos_mean_mae": 58.19400664292878
118
+ },
119
+ "per_repo_worst10": [
120
+ {
121
+ "repo": "Cidoyi/so100_all_notes",
122
+ "elements": 215,
123
+ "mae": 121.31421661376953,
124
+ "floor_mae": 42.90715408325195,
125
+ "rewear_repo_mae": 3.015285015106201,
126
+ "truth_outside_frac": 0.8333333333333334
127
+ },
128
+ {
129
+ "repo": "Cidoyi/so100_all_notes_3",
130
+ "elements": 354,
131
+ "mae": 115.09871673583984,
132
+ "floor_mae": 38.250736236572266,
133
+ "rewear_repo_mae": 8.197112083435059,
134
+ "truth_outside_frac": 0.8333333333333334
135
+ },
136
+ {
137
+ "repo": "roboticshack/team11_pianobot",
138
+ "elements": 452,
139
+ "mae": 107.51007080078125,
140
+ "floor_mae": 35.70561981201172,
141
+ "rewear_repo_mae": 15.485651016235352,
142
+ "truth_outside_frac": 0.7901917404129793
143
+ },
144
+ {
145
+ "repo": "Cidoyi/so100_all_notes_1",
146
+ "elements": 389,
147
+ "mae": 107.12379455566406,
148
+ "floor_mae": 34.05473327636719,
149
+ "rewear_repo_mae": 15.588462829589844,
150
+ "truth_outside_frac": 0.786203941730934
151
+ },
152
+ {
153
+ "repo": "kantine/flip_A2",
154
+ "elements": 120,
155
+ "mae": 101.39452362060547,
156
+ "floor_mae": 55.31953048706055,
157
+ "rewear_repo_mae": 26.316650390625,
158
+ "truth_outside_frac": 0.8333333333333334
159
+ },
160
+ {
161
+ "repo": "siqiLi/so100_test16",
162
+ "elements": 548,
163
+ "mae": 96.68895721435547,
164
+ "floor_mae": 32.16460418701172,
165
+ "rewear_repo_mae": 41.742645263671875,
166
+ "truth_outside_frac": 0.40206812652068125
167
+ },
168
+ {
169
+ "repo": "juni3227/so100_test_cap02_sideway",
170
+ "elements": 720,
171
+ "mae": 96.2467269897461,
172
+ "floor_mae": 28.773996353149414,
173
+ "rewear_repo_mae": 33.46946716308594,
174
+ "truth_outside_frac": 0.47824074074074074
175
+ },
176
+ {
177
+ "repo": "shuohsuan/grasp_data_tg_10",
178
+ "elements": 115,
179
+ "mae": 95.52738952636719,
180
+ "floor_mae": 29.273643493652344,
181
+ "rewear_repo_mae": 16.6530704498291,
182
+ "truth_outside_frac": 0.3289855072463768
183
+ },
184
+ {
185
+ "repo": "kantine/flip_A3",
186
+ "elements": 120,
187
+ "mae": 93.62525939941406,
188
+ "floor_mae": 41.843135833740234,
189
+ "rewear_repo_mae": 25.944631576538086,
190
+ "truth_outside_frac": 0.7041666666666667
191
+ },
192
+ {
193
+ "repo": "bjb7/so101_pen_mug_10_13",
194
+ "elements": 120,
195
+ "mae": 93.45197296142578,
196
+ "floor_mae": 27.867626190185547,
197
+ "rewear_repo_mae": 17.589902877807617,
198
+ "truth_outside_frac": 0.4777777777777778
199
+ }
200
+ ]
201
+ }