{ "npz": "reports/eval__grasp_sft_v2_demosonly_1gpu_disc__step_001000__panel_v2_k4l2_euler10_draws1_stable.npz", "checkpoint": "/home/ubuntu/checkpoints/finetune/grasp_sft_v2_demosonly_1gpu_disc/step_001000", "wear_facts": { "leg_molmo_norm": "checkpoint", "flow_normalization_tag": "q01q99", "train_per_dataset_flow_norm": false, "per_dataset_stats_keys": [ "grasp_demos_v2/merged" ], "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)", "worn_table": { "mean": [ 0.09233876317739487, -7.418671607971191, 22.1364803314209, 51.24595642089844, 82.55037689208984, 12.741997718811035 ], "std": [ 27.988689422607422, 40.790802001953125, 31.966609954833984, 20.678668975830078, 77.26783752441406, 19.013696670532227 ], "q01": [ -54.11045837402344, 55.71287155151367, -65.48938751220703, 11.65783405303955, -157.21087646484375, 0.0 ], "q99": [ 50.82916259765625, -113.50212097167969, 97.0, 95.00440979003906, 157.21087646484375, 41.692020416259766 ] }, "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" }, "caveats": [ "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", "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" ], "anchors_reproduced": { "bijou@1000": 58.143384213383385, "state-copy": 8.367757873526733 }, "box_audit": { "truth_outside_frac_per_joint": { "shoulder_pan": 0.07369556616706689, "shoulder_lift": 0.7296525803453955, "elbow_flex": 0.5511537718402385, "wrist_flex": 0.19014923587917598, "wrist_roll": 0.027241068389043766, "gripper": 0.26894430796691415 }, "truth_any_joint_outside_frac": 0.8583248226558834, "floor_chunk_mae": 14.401469230651855, "floor_per_joint": { "shoulder_pan": 1.4592567682266235, "shoulder_lift": 52.04825210571289, "elbow_flex": 24.017120361328125, "wrist_flex": 6.615574359893799, "wrist_roll": 0.5258612036705017, "gripper": 1.74805748462677 }, "pred_edge_saturation_per_joint": { "shoulder_pan": 0.000588846872211028, "shoulder_lift": 0.0014905186452841646, "elbow_flex": 0.00014491153495818266, "wrist_flex": 0.002872928685123336, "wrist_roll": 0.0016561318280935162, "gripper": 0.11585562210751976 } }, "actual_per_joint_mae": { "shoulder_pan": 23.462173461914062, "shoulder_lift": 104.43495178222656, "elbow_flex": 99.3005599975586, "wrist_flex": 27.106822967529297, "wrist_roll": 71.17012023925781, "gripper": 23.381319046020508 }, "rewear": { "round_trip_worst_abs_deg": 1.52587890625e-05, "repo_rows_chunk_mae": 27.395334001919885, "repo_rows_per_joint": { "shoulder_pan": 21.10439109802246, "shoulder_lift": 41.8028678894043, "elbow_flex": 36.7511100769043, "wrist_flex": 22.337148666381836, "wrist_roll": 22.218639373779297, "gripper": 20.153600692749023 }, "released_table_chunk_mae": 54.399765688015435, "released_table_per_joint": { "shoulder_pan": 24.25238609313965, "shoulder_lift": 105.2990951538086, "elbow_flex": 98.5770263671875, "wrist_flex": 27.263837814331055, "wrist_roll": 46.14845275878906, "gripper": 24.851577758789062 }, "null_repo_midpoint_chunk_mae": 25.154475696265422, "n_repo_rows": 838 }, "demos_prior_collapse": { "pred_vs_demos_mean_mae": 22.636736684240066, "truth_vs_demos_mean_mae": 58.19400664292878 }, "per_repo_worst10": [ { "repo": "Cidoyi/so100_all_notes", "elements": 215, "mae": 121.31421661376953, "floor_mae": 42.90715408325195, "rewear_repo_mae": 3.015285015106201, "truth_outside_frac": 0.8333333333333334 }, { "repo": "Cidoyi/so100_all_notes_3", "elements": 354, "mae": 115.09871673583984, "floor_mae": 38.250736236572266, "rewear_repo_mae": 8.197112083435059, "truth_outside_frac": 0.8333333333333334 }, { "repo": "roboticshack/team11_pianobot", "elements": 452, "mae": 107.51007080078125, "floor_mae": 35.70561981201172, "rewear_repo_mae": 15.485651016235352, "truth_outside_frac": 0.7901917404129793 }, { "repo": "Cidoyi/so100_all_notes_1", "elements": 389, "mae": 107.12379455566406, "floor_mae": 34.05473327636719, "rewear_repo_mae": 15.588462829589844, "truth_outside_frac": 0.786203941730934 }, { "repo": "kantine/flip_A2", "elements": 120, "mae": 101.39452362060547, "floor_mae": 55.31953048706055, "rewear_repo_mae": 26.316650390625, "truth_outside_frac": 0.8333333333333334 }, { "repo": "siqiLi/so100_test16", "elements": 548, "mae": 96.68895721435547, "floor_mae": 32.16460418701172, "rewear_repo_mae": 41.742645263671875, "truth_outside_frac": 0.40206812652068125 }, { "repo": "juni3227/so100_test_cap02_sideway", "elements": 720, "mae": 96.2467269897461, "floor_mae": 28.773996353149414, "rewear_repo_mae": 33.46946716308594, "truth_outside_frac": 0.47824074074074074 }, { "repo": "shuohsuan/grasp_data_tg_10", "elements": 115, "mae": 95.52738952636719, "floor_mae": 29.273643493652344, "rewear_repo_mae": 16.6530704498291, "truth_outside_frac": 0.3289855072463768 }, { "repo": "kantine/flip_A3", "elements": 120, "mae": 93.62525939941406, "floor_mae": 41.843135833740234, "rewear_repo_mae": 25.944631576538086, "truth_outside_frac": 0.7041666666666667 }, { "repo": "bjb7/so101_pen_mug_10_13", "elements": 120, "mae": 93.45197296142578, "floor_mae": 27.867626190185547, "rewear_repo_mae": 17.589902877807617, "truth_outside_frac": 0.4777777777777778 } ] }