Download continual_metrics.json from tml-lab/split-itail-3000: direct link, hf CLI and curl.
- Browser
- Download file 31.9 kB
-
https://huggingface.co/tml-lab/split-itail-3000/resolve/main/continual_metrics.json
- Command line
-
hf download hf://tml-lab/split-itail-3000/continual_metrics.json
-
curl -L -o continual_metrics.json https://huggingface.co/tml-lab/split-itail-3000/resolve/main/continual_metrics.json
31.9 kB
| { | |
| "schema_version": 1, | |
| "protocol": "split-sequential-full-horizon", | |
| "generated_at": "2026-09-25T19:21:19+09:00", | |
| "checkpoint_selection": "last-per-split", | |
| "units": "success rates are fractions in [0, 1]; NBT is a signed difference", | |
| "indexing": "stage_index, split_index, and task positions are zero-based", | |
| "definitions": { | |
| "sr": "mean task success immediately after each task's last split visit", | |
| "final_sr": "alias of sr: last-visit acquisition, not single-final-policy success", | |
| "fwt": "mean acquisition success over all visits; endpoint proxy, not learning-curve FWT", | |
| "nbt": "mean over visits of mean acquisition minus same-task success at ALL later checkpoints, including revisits; last visit contributes zero; negatives retained", | |
| "normalized_nbt": "mean defined visit NBT/acquisition ratios; zero-acquisition and final visits excluded", | |
| "final_policy_sr": "mean task success evaluated at the single final checkpoint" | |
| }, | |
| "complete_sequences": 4, | |
| "discovered_sequences": 4, | |
| "incomplete_sequences": [], | |
| "sequences": [ | |
| { | |
| "sequence_id": "10-26-07_smolvla_instruction_tail_libero_10_2e-3_split2/libero_10", | |
| "task_order": [ | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9 | |
| ], | |
| "visit_task_order": [ | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9, | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9 | |
| ], | |
| "checkpoints": [ | |
| "libero_10_task00_005000", | |
| "libero_10_task01_010000", | |
| "libero_10_task02_015000", | |
| "libero_10_task03_020000", | |
| "libero_10_task04_025000", | |
| "libero_10_task05_030000", | |
| "libero_10_task06_035000", | |
| "libero_10_task07_040000", | |
| "libero_10_task08_045000", | |
| "libero_10_task09_050000", | |
| "libero_10_task00_055000", | |
| "libero_10_task01_060000", | |
| "libero_10_task02_065000", | |
| "libero_10_task03_070000", | |
| "libero_10_task04_075000", | |
| "libero_10_task05_080000", | |
| "libero_10_task06_085000", | |
| "libero_10_task07_090000", | |
| "libero_10_task08_095000", | |
| "libero_10_task09_100000" | |
| ], | |
| "run": "10-26-07_smolvla_instruction_tail_libero_10_2e-3_split2", | |
| "suite": "libero_10", | |
| "num_tasks": 10, | |
| "split": 2, | |
| "num_visits": 20, | |
| "all_visits_complete": true, | |
| "sr": 0.396, | |
| "final_sr": 0.396, | |
| "fwt": 0.398, | |
| "nbt": -0.00014342426149701736, | |
| "normalized_nbt": -0.04263810704068235, | |
| "final_policy_sr": 0.396, | |
| "sr_by_split": [ | |
| 0.4, | |
| 0.396 | |
| ], | |
| "last_visit_sr_by_task": [ | |
| 0.28, | |
| 0.34, | |
| 0.7, | |
| 0.8, | |
| 0.28, | |
| 0.72, | |
| 0.1, | |
| 0.18, | |
| 0.22, | |
| 0.34 | |
| ], | |
| "acquisition_sr_by_visit": [ | |
| 0.2, | |
| 0.16, | |
| 0.9, | |
| 0.86, | |
| 0.16, | |
| 0.78, | |
| 0.24, | |
| 0.14, | |
| 0.16, | |
| 0.4, | |
| 0.28, | |
| 0.34, | |
| 0.7, | |
| 0.8, | |
| 0.28, | |
| 0.72, | |
| 0.1, | |
| 0.18, | |
| 0.22, | |
| 0.34 | |
| ], | |
| "nbt_by_visit": [ | |
| -0.04210526315789474, | |
| -0.09000000000000001, | |
| 0.09411764705882356, | |
| 0.026249999999999975, | |
| -0.048000000000000015, | |
| 0.021428571428571446, | |
| 0.043076923076923075, | |
| -0.009999999999999995, | |
| -0.003636363636363636, | |
| 0.006, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0 | |
| ], | |
| "normalized_nbt_by_visit": [ | |
| -0.2105263157894737, | |
| -0.5625, | |
| 0.10457516339869284, | |
| 0.03052325581395346, | |
| -0.3000000000000001, | |
| 0.027472527472527493, | |
| 0.1794871794871795, | |
| -0.07142857142857138, | |
| -0.022727272727272724, | |
| 0.015, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| null | |
| ], | |
| "stage_metrics": [ | |
| { | |
| "stage_index": 0, | |
| "stage_sr": 0.2, | |
| "acquisition_sr": 0.2, | |
| "nbt": -0.04210526315789474, | |
| "normalized_nbt": -0.2105263157894737, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 1, | |
| "stage_sr": 0.18, | |
| "acquisition_sr": 0.16, | |
| "nbt": -0.09000000000000001, | |
| "normalized_nbt": -0.5625, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 2, | |
| "stage_sr": 0.42, | |
| "acquisition_sr": 0.9, | |
| "nbt": 0.09411764705882356, | |
| "normalized_nbt": 0.10457516339869284, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 3, | |
| "stage_sr": 0.53, | |
| "acquisition_sr": 0.86, | |
| "nbt": 0.026249999999999975, | |
| "normalized_nbt": 0.03052325581395346, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 4, | |
| "stage_sr": 0.45600000000000007, | |
| "acquisition_sr": 0.16, | |
| "nbt": -0.048000000000000015, | |
| "normalized_nbt": -0.3000000000000001, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 5, | |
| "stage_sr": 0.51, | |
| "acquisition_sr": 0.78, | |
| "nbt": 0.021428571428571446, | |
| "normalized_nbt": 0.027472527472527493, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 6, | |
| "stage_sr": 0.4714285714285714, | |
| "acquisition_sr": 0.24, | |
| "nbt": 0.043076923076923075, | |
| "normalized_nbt": 0.1794871794871795, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 7, | |
| "stage_sr": 0.43, | |
| "acquisition_sr": 0.14, | |
| "nbt": -0.009999999999999995, | |
| "normalized_nbt": -0.07142857142857138, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 8, | |
| "stage_sr": 0.4, | |
| "acquisition_sr": 0.16, | |
| "nbt": -0.003636363636363636, | |
| "normalized_nbt": -0.022727272727272724, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 9, | |
| "stage_sr": 0.4, | |
| "acquisition_sr": 0.4, | |
| "nbt": 0.006, | |
| "normalized_nbt": 0.015, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 10, | |
| "stage_sr": 0.40800000000000003, | |
| "acquisition_sr": 0.28, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0007272727272727274 | |
| }, | |
| { | |
| "stage_index": 11, | |
| "stage_sr": 0.426, | |
| "acquisition_sr": 0.34, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0027121212121212126 | |
| }, | |
| { | |
| "stage_index": 12, | |
| "stage_sr": 0.406, | |
| "acquisition_sr": 0.7, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0025174825174825175 | |
| }, | |
| { | |
| "stage_index": 13, | |
| "stage_sr": 0.392, | |
| "acquisition_sr": 0.8, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0008036963036963038 | |
| }, | |
| { | |
| "stage_index": 14, | |
| "stage_sr": 0.41200000000000003, | |
| "acquisition_sr": 0.28, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.001447574647574648 | |
| }, | |
| { | |
| "stage_index": 15, | |
| "stage_sr": 0.406, | |
| "acquisition_sr": 0.72, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0015085539460539467 | |
| }, | |
| { | |
| "stage_index": 16, | |
| "stage_sr": 0.392, | |
| "acquisition_sr": 0.1, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0007176940706352467 | |
| }, | |
| { | |
| "stage_index": 17, | |
| "stage_sr": 0.396, | |
| "acquisition_sr": 0.18, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.00036378499858892094 | |
| }, | |
| { | |
| "stage_index": 18, | |
| "stage_sr": 0.402, | |
| "acquisition_sr": 0.22, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.00041443231690909766 | |
| }, | |
| { | |
| "stage_index": 19, | |
| "stage_sr": 0.396, | |
| "acquisition_sr": 0.34, | |
| "nbt": 0.0, | |
| "normalized_nbt": null, | |
| "cumulative_nbt": -0.00014342426149701736 | |
| } | |
| ] | |
| }, | |
| { | |
| "sequence_id": "10-26-08_smolvla_instruction_tail_libero_goal_2e-3_split2/libero_goal", | |
| "task_order": [ | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9 | |
| ], | |
| "visit_task_order": [ | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9, | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9 | |
| ], | |
| "checkpoints": [ | |
| "libero_goal_task00_005000", | |
| "libero_goal_task01_010000", | |
| "libero_goal_task02_015000", | |
| "libero_goal_task03_020000", | |
| "libero_goal_task04_025000", | |
| "libero_goal_task05_030000", | |
| "libero_goal_task06_035000", | |
| "libero_goal_task07_040000", | |
| "libero_goal_task08_045000", | |
| "libero_goal_task09_050000", | |
| "libero_goal_task00_055000", | |
| "libero_goal_task01_060000", | |
| "libero_goal_task02_065000", | |
| "libero_goal_task03_070000", | |
| "libero_goal_task04_075000", | |
| "libero_goal_task05_080000", | |
| "libero_goal_task06_085000", | |
| "libero_goal_task07_090000", | |
| "libero_goal_task08_095000", | |
| "libero_goal_task09_100000" | |
| ], | |
| "run": "10-26-08_smolvla_instruction_tail_libero_goal_2e-3_split2", | |
| "suite": "libero_goal", | |
| "num_tasks": 10, | |
| "split": 2, | |
| "num_visits": 20, | |
| "all_visits_complete": true, | |
| "sr": 0.82, | |
| "final_sr": 0.82, | |
| "fwt": 0.828, | |
| "nbt": 0.005954152078571586, | |
| "normalized_nbt": 0.008376547423910453, | |
| "final_policy_sr": 0.82, | |
| "sr_by_split": [ | |
| 0.836, | |
| 0.82 | |
| ], | |
| "last_visit_sr_by_task": [ | |
| 0.94, | |
| 0.98, | |
| 0.84, | |
| 0.54, | |
| 0.86, | |
| 0.94, | |
| 0.46, | |
| 1.0, | |
| 0.98, | |
| 0.66 | |
| ], | |
| "acquisition_sr_by_visit": [ | |
| 0.9, | |
| 0.96, | |
| 0.96, | |
| 0.64, | |
| 0.9, | |
| 0.98, | |
| 0.66, | |
| 0.94, | |
| 0.92, | |
| 0.5, | |
| 0.94, | |
| 0.98, | |
| 0.84, | |
| 0.54, | |
| 0.86, | |
| 0.94, | |
| 0.46, | |
| 1.0, | |
| 0.98, | |
| 0.66 | |
| ], | |
| "nbt_by_visit": [ | |
| -0.02105263157894733, | |
| -0.010000000000000009, | |
| 0.05647058823529411, | |
| 0.04374999999999999, | |
| 0.016000000000000014, | |
| 0.0142857142857143, | |
| 0.06153846153846154, | |
| -0.015000000000000013, | |
| -0.010909090909090898, | |
| -0.016000000000000004, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0 | |
| ], | |
| "normalized_nbt_by_visit": [ | |
| -0.023391812865497033, | |
| -0.010416666666666676, | |
| 0.058823529411764705, | |
| 0.06835937499999999, | |
| 0.01777777777777779, | |
| 0.014577259475218672, | |
| 0.09324009324009325, | |
| -0.015957446808510654, | |
| -0.01185770750988141, | |
| -0.03200000000000001, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| null | |
| ], | |
| "stage_metrics": [ | |
| { | |
| "stage_index": 0, | |
| "stage_sr": 0.9, | |
| "acquisition_sr": 0.9, | |
| "nbt": -0.02105263157894733, | |
| "normalized_nbt": -0.023391812865497033, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 1, | |
| "stage_sr": 0.9299999999999999, | |
| "acquisition_sr": 0.96, | |
| "nbt": -0.010000000000000009, | |
| "normalized_nbt": -0.010416666666666676, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 2, | |
| "stage_sr": 0.94, | |
| "acquisition_sr": 0.96, | |
| "nbt": 0.05647058823529411, | |
| "normalized_nbt": 0.058823529411764705, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 3, | |
| "stage_sr": 0.865, | |
| "acquisition_sr": 0.64, | |
| "nbt": 0.04374999999999999, | |
| "normalized_nbt": 0.06835937499999999, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 4, | |
| "stage_sr": 0.8720000000000001, | |
| "acquisition_sr": 0.9, | |
| "nbt": 0.016000000000000014, | |
| "normalized_nbt": 0.01777777777777779, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 5, | |
| "stage_sr": 0.89, | |
| "acquisition_sr": 0.98, | |
| "nbt": 0.0142857142857143, | |
| "normalized_nbt": 0.014577259475218672, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 6, | |
| "stage_sr": 0.8571428571428571, | |
| "acquisition_sr": 0.66, | |
| "nbt": 0.06153846153846154, | |
| "normalized_nbt": 0.09324009324009325, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 7, | |
| "stage_sr": 0.8674999999999999, | |
| "acquisition_sr": 0.94, | |
| "nbt": -0.015000000000000013, | |
| "normalized_nbt": -0.015957446808510654, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 8, | |
| "stage_sr": 0.8733333333333334, | |
| "acquisition_sr": 0.92, | |
| "nbt": -0.010909090909090898, | |
| "normalized_nbt": -0.01185770750988141, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 9, | |
| "stage_sr": 0.836, | |
| "acquisition_sr": 0.5, | |
| "nbt": -0.016000000000000004, | |
| "normalized_nbt": -0.03200000000000001, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 10, | |
| "stage_sr": 0.8400000000000001, | |
| "acquisition_sr": 0.94, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0003636363636363629 | |
| }, | |
| { | |
| "stage_index": 11, | |
| "stage_sr": 0.842, | |
| "acquisition_sr": 0.98, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0007727272727272717 | |
| }, | |
| { | |
| "stage_index": 12, | |
| "stage_sr": 0.8300000000000001, | |
| "acquisition_sr": 0.84, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.00012587412587412465 | |
| }, | |
| { | |
| "stage_index": 13, | |
| "stage_sr": 0.82, | |
| "acquisition_sr": 0.54, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.0010364635364635375 | |
| }, | |
| { | |
| "stage_index": 14, | |
| "stage_sr": 0.8160000000000001, | |
| "acquisition_sr": 0.86, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.0021161505161505175 | |
| }, | |
| { | |
| "stage_index": 15, | |
| "stage_sr": 0.8119999999999999, | |
| "acquisition_sr": 0.94, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.0031283091908091924 | |
| }, | |
| { | |
| "stage_index": 16, | |
| "stage_sr": 0.792, | |
| "acquisition_sr": 0.46, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.005023476523476526 | |
| }, | |
| { | |
| "stage_index": 17, | |
| "stage_sr": 0.7979999999999999, | |
| "acquisition_sr": 1.0, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.00604511583188054 | |
| }, | |
| { | |
| "stage_index": 18, | |
| "stage_sr": 0.8039999999999999, | |
| "acquisition_sr": 0.98, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.006461713209391231 | |
| }, | |
| { | |
| "stage_index": 19, | |
| "stage_sr": 0.82, | |
| "acquisition_sr": 0.66, | |
| "nbt": 0.0, | |
| "normalized_nbt": null, | |
| "cumulative_nbt": 0.005954152078571586 | |
| } | |
| ] | |
| }, | |
| { | |
| "sequence_id": "10-26-10_smolvla_instruction_tail_libero_object_2e-3_split2/libero_object", | |
| "task_order": [ | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9 | |
| ], | |
| "visit_task_order": [ | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9, | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9 | |
| ], | |
| "checkpoints": [ | |
| "libero_object_task00_005000", | |
| "libero_object_task01_010000", | |
| "libero_object_task02_015000", | |
| "libero_object_task03_020000", | |
| "libero_object_task04_025000", | |
| "libero_object_task05_030000", | |
| "libero_object_task06_035000", | |
| "libero_object_task07_040000", | |
| "libero_object_task08_045000", | |
| "libero_object_task09_050000", | |
| "libero_object_task00_055000", | |
| "libero_object_task01_060000", | |
| "libero_object_task02_065000", | |
| "libero_object_task03_070000", | |
| "libero_object_task04_075000", | |
| "libero_object_task05_080000", | |
| "libero_object_task06_085000", | |
| "libero_object_task07_090000", | |
| "libero_object_task08_095000", | |
| "libero_object_task09_100000" | |
| ], | |
| "run": "10-26-10_smolvla_instruction_tail_libero_object_2e-3_split2", | |
| "suite": "libero_object", | |
| "num_tasks": 10, | |
| "split": 2, | |
| "num_visits": 20, | |
| "all_visits_complete": true, | |
| "sr": 0.818, | |
| "final_sr": 0.818, | |
| "fwt": 0.767, | |
| "nbt": -0.014132960963185425, | |
| "normalized_nbt": -0.02890769180838432, | |
| "final_policy_sr": 0.818, | |
| "sr_by_split": [ | |
| 0.716, | |
| 0.818 | |
| ], | |
| "last_visit_sr_by_task": [ | |
| 0.52, | |
| 0.68, | |
| 0.92, | |
| 0.78, | |
| 0.9, | |
| 0.8, | |
| 0.9, | |
| 0.9, | |
| 0.98, | |
| 0.8 | |
| ], | |
| "acquisition_sr_by_visit": [ | |
| 0.54, | |
| 0.78, | |
| 0.98, | |
| 0.48, | |
| 0.94, | |
| 0.64, | |
| 0.66, | |
| 0.6, | |
| 0.66, | |
| 0.88, | |
| 0.52, | |
| 0.68, | |
| 0.92, | |
| 0.78, | |
| 0.9, | |
| 0.8, | |
| 0.9, | |
| 0.9, | |
| 0.98, | |
| 0.8 | |
| ], | |
| "nbt_by_visit": [ | |
| 0.010526315789473694, | |
| 0.04999999999999999, | |
| 0.028235294117647032, | |
| -0.13125000000000003, | |
| 0.01599999999999997, | |
| -0.057142857142857155, | |
| -0.07384615384615384, | |
| -0.07500000000000001, | |
| -0.058181818181818175, | |
| 0.007999999999999997, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0 | |
| ], | |
| "normalized_nbt_by_visit": [ | |
| 0.019493177387914246, | |
| 0.06410256410256408, | |
| 0.02881152460984391, | |
| -0.27343750000000006, | |
| 0.01702127659574465, | |
| -0.0892857142857143, | |
| -0.11188811188811187, | |
| -0.12500000000000003, | |
| -0.08815426997245178, | |
| 0.009090909090909087, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| null | |
| ], | |
| "stage_metrics": [ | |
| { | |
| "stage_index": 0, | |
| "stage_sr": 0.54, | |
| "acquisition_sr": 0.54, | |
| "nbt": 0.010526315789473694, | |
| "normalized_nbt": 0.019493177387914246, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 1, | |
| "stage_sr": 0.66, | |
| "acquisition_sr": 0.78, | |
| "nbt": 0.04999999999999999, | |
| "normalized_nbt": 0.06410256410256408, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 2, | |
| "stage_sr": 0.7666666666666666, | |
| "acquisition_sr": 0.98, | |
| "nbt": 0.028235294117647032, | |
| "normalized_nbt": 0.02881152460984391, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 3, | |
| "stage_sr": 0.6950000000000001, | |
| "acquisition_sr": 0.48, | |
| "nbt": -0.13125000000000003, | |
| "normalized_nbt": -0.27343750000000006, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 4, | |
| "stage_sr": 0.744, | |
| "acquisition_sr": 0.94, | |
| "nbt": 0.01599999999999997, | |
| "normalized_nbt": 0.01702127659574465, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 5, | |
| "stage_sr": 0.7266666666666667, | |
| "acquisition_sr": 0.64, | |
| "nbt": -0.057142857142857155, | |
| "normalized_nbt": -0.0892857142857143, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 6, | |
| "stage_sr": 0.7171428571428571, | |
| "acquisition_sr": 0.66, | |
| "nbt": -0.07384615384615384, | |
| "normalized_nbt": -0.11188811188811187, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 7, | |
| "stage_sr": 0.7025, | |
| "acquisition_sr": 0.6, | |
| "nbt": -0.07500000000000001, | |
| "normalized_nbt": -0.12500000000000003, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 8, | |
| "stage_sr": 0.6977777777777778, | |
| "acquisition_sr": 0.66, | |
| "nbt": -0.058181818181818175, | |
| "normalized_nbt": -0.08815426997245178, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 9, | |
| "stage_sr": 0.716, | |
| "acquisition_sr": 0.88, | |
| "nbt": 0.007999999999999997, | |
| "normalized_nbt": 0.009090909090909087, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 10, | |
| "stage_sr": 0.714, | |
| "acquisition_sr": 0.52, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.00018181818181818197 | |
| }, | |
| { | |
| "stage_index": 11, | |
| "stage_sr": 0.704, | |
| "acquisition_sr": 0.68, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.0011363636363636365 | |
| }, | |
| { | |
| "stage_index": 12, | |
| "stage_sr": 0.6980000000000001, | |
| "acquisition_sr": 0.92, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.0022447552447552445 | |
| }, | |
| { | |
| "stage_index": 13, | |
| "stage_sr": 0.728, | |
| "acquisition_sr": 0.78, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.0008616383616383604 | |
| }, | |
| { | |
| "stage_index": 14, | |
| "stage_sr": 0.724, | |
| "acquisition_sr": 0.9, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.00015777555777555564 | |
| }, | |
| { | |
| "stage_index": 15, | |
| "stage_sr": 0.74, | |
| "acquisition_sr": 0.8, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0013469655344655377 | |
| }, | |
| { | |
| "stage_index": 16, | |
| "stage_sr": 0.764, | |
| "acquisition_sr": 0.9, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0038364723511782376 | |
| }, | |
| { | |
| "stage_index": 17, | |
| "stage_sr": 0.794, | |
| "acquisition_sr": 0.9, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.007294986875869234 | |
| }, | |
| { | |
| "stage_index": 18, | |
| "stage_sr": 0.826, | |
| "acquisition_sr": 0.98, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.011483246629531462 | |
| }, | |
| { | |
| "stage_index": 19, | |
| "stage_sr": 0.818, | |
| "acquisition_sr": 0.8, | |
| "nbt": 0.0, | |
| "normalized_nbt": null, | |
| "cumulative_nbt": -0.014132960963185425 | |
| } | |
| ] | |
| }, | |
| { | |
| "sequence_id": "10-26-12_smolvla_instruction_tail_libero_spatial_2e-3_split2/libero_spatial", | |
| "task_order": [ | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9 | |
| ], | |
| "visit_task_order": [ | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9, | |
| 0, | |
| 1, | |
| 2, | |
| 3, | |
| 4, | |
| 5, | |
| 6, | |
| 7, | |
| 8, | |
| 9 | |
| ], | |
| "checkpoints": [ | |
| "libero_spatial_task00_005000", | |
| "libero_spatial_task01_010000", | |
| "libero_spatial_task02_015000", | |
| "libero_spatial_task03_020000", | |
| "libero_spatial_task04_025000", | |
| "libero_spatial_task05_030000", | |
| "libero_spatial_task06_035000", | |
| "libero_spatial_task07_040000", | |
| "libero_spatial_task08_045000", | |
| "libero_spatial_task09_050000", | |
| "libero_spatial_task00_055000", | |
| "libero_spatial_task01_060000", | |
| "libero_spatial_task02_065000", | |
| "libero_spatial_task03_070000", | |
| "libero_spatial_task04_075000", | |
| "libero_spatial_task05_080000", | |
| "libero_spatial_task06_085000", | |
| "libero_spatial_task07_090000", | |
| "libero_spatial_task08_095000", | |
| "libero_spatial_task09_100000" | |
| ], | |
| "run": "10-26-12_smolvla_instruction_tail_libero_spatial_2e-3_split2", | |
| "suite": "libero_spatial", | |
| "num_tasks": 10, | |
| "split": 2, | |
| "num_visits": 20, | |
| "all_visits_complete": true, | |
| "sr": 0.714, | |
| "final_sr": 0.714, | |
| "fwt": 0.749, | |
| "nbt": 0.008701261509388446, | |
| "normalized_nbt": 0.00897760067708685, | |
| "final_policy_sr": 0.714, | |
| "sr_by_split": [ | |
| 0.784, | |
| 0.714 | |
| ], | |
| "last_visit_sr_by_task": [ | |
| 0.96, | |
| 0.98, | |
| 0.9, | |
| 0.7, | |
| 0.22, | |
| 0.52, | |
| 0.92, | |
| 0.8, | |
| 0.6, | |
| 0.54 | |
| ], | |
| "acquisition_sr_by_visit": [ | |
| 0.76, | |
| 0.84, | |
| 0.92, | |
| 0.98, | |
| 0.74, | |
| 0.42, | |
| 0.92, | |
| 0.8, | |
| 0.76, | |
| 0.7, | |
| 0.96, | |
| 0.98, | |
| 0.9, | |
| 0.7, | |
| 0.22, | |
| 0.52, | |
| 0.92, | |
| 0.8, | |
| 0.6, | |
| 0.54 | |
| ], | |
| "nbt_by_visit": [ | |
| -0.10526315789473682, | |
| -0.07, | |
| 0.00941176470588236, | |
| 0.12250000000000001, | |
| 0.20800000000000002, | |
| -0.03571428571428573, | |
| 0.0, | |
| 0.0, | |
| 0.029090909090909098, | |
| 0.015999999999999993, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0 | |
| ], | |
| "normalized_nbt_by_visit": [ | |
| -0.13850415512465372, | |
| -0.08333333333333334, | |
| 0.010230179028133, | |
| 0.12500000000000003, | |
| 0.2810810810810811, | |
| -0.08503401360544223, | |
| 0.0, | |
| 0.0, | |
| 0.038277511961722493, | |
| 0.02285714285714285, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| 0.0, | |
| null | |
| ], | |
| "stage_metrics": [ | |
| { | |
| "stage_index": 0, | |
| "stage_sr": 0.76, | |
| "acquisition_sr": 0.76, | |
| "nbt": -0.10526315789473682, | |
| "normalized_nbt": -0.13850415512465372, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 1, | |
| "stage_sr": 0.8, | |
| "acquisition_sr": 0.84, | |
| "nbt": -0.07, | |
| "normalized_nbt": -0.08333333333333334, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 2, | |
| "stage_sr": 0.84, | |
| "acquisition_sr": 0.92, | |
| "nbt": 0.00941176470588236, | |
| "normalized_nbt": 0.010230179028133, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 3, | |
| "stage_sr": 0.875, | |
| "acquisition_sr": 0.98, | |
| "nbt": 0.12250000000000001, | |
| "normalized_nbt": 0.12500000000000003, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 4, | |
| "stage_sr": 0.8480000000000001, | |
| "acquisition_sr": 0.74, | |
| "nbt": 0.20800000000000002, | |
| "normalized_nbt": 0.2810810810810811, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 5, | |
| "stage_sr": 0.7766666666666667, | |
| "acquisition_sr": 0.42, | |
| "nbt": -0.03571428571428573, | |
| "normalized_nbt": -0.08503401360544223, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 6, | |
| "stage_sr": 0.7971428571428572, | |
| "acquisition_sr": 0.92, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 7, | |
| "stage_sr": 0.7975, | |
| "acquisition_sr": 0.8, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 8, | |
| "stage_sr": 0.7933333333333333, | |
| "acquisition_sr": 0.76, | |
| "nbt": 0.029090909090909098, | |
| "normalized_nbt": 0.038277511961722493, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 9, | |
| "stage_sr": 0.784, | |
| "acquisition_sr": 0.7, | |
| "nbt": 0.015999999999999993, | |
| "normalized_nbt": 0.02285714285714285, | |
| "cumulative_nbt": 0.0 | |
| }, | |
| { | |
| "stage_index": 10, | |
| "stage_sr": 0.8039999999999999, | |
| "acquisition_sr": 0.96, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.001818181818181818 | |
| }, | |
| { | |
| "stage_index": 11, | |
| "stage_sr": 0.818, | |
| "acquisition_sr": 0.98, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.004196969696969696 | |
| }, | |
| { | |
| "stage_index": 12, | |
| "stage_sr": 0.8160000000000001, | |
| "acquisition_sr": 0.9, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.005650349650349649 | |
| }, | |
| { | |
| "stage_index": 13, | |
| "stage_sr": 0.788, | |
| "acquisition_sr": 0.7, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.004635864135864135 | |
| }, | |
| { | |
| "stage_index": 14, | |
| "stage_sr": 0.736, | |
| "acquisition_sr": 0.22, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": -0.0004397602397602382 | |
| }, | |
| { | |
| "stage_index": 15, | |
| "stage_sr": 0.746, | |
| "acquisition_sr": 0.52, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.0019187062937062946 | |
| }, | |
| { | |
| "stage_index": 16, | |
| "stage_sr": 0.746, | |
| "acquisition_sr": 0.92, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.0036244050067579495 | |
| }, | |
| { | |
| "stage_index": 17, | |
| "stage_sr": 0.746, | |
| "acquisition_sr": 0.8, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.004868464052287583 | |
| }, | |
| { | |
| "stage_index": 18, | |
| "stage_sr": 0.73, | |
| "acquisition_sr": 0.6, | |
| "nbt": 0.0, | |
| "normalized_nbt": 0.0, | |
| "cumulative_nbt": 0.006621721158098869 | |
| }, | |
| { | |
| "stage_index": 19, | |
| "stage_sr": 0.714, | |
| "acquisition_sr": 0.54, | |
| "nbt": 0.0, | |
| "normalized_nbt": null, | |
| "cumulative_nbt": 0.008701261509388446 | |
| } | |
| ] | |
| } | |
| ] | |
| } | |