RealArDor's picture
download
raw
19.3 kB
{
"best_step": 300,
"best_dev_selection_ce": 0.768348061442913,
"selected_on": "dev only",
"stage": "sft",
"loss": "ce",
"completed_steps": 300,
"metrics_by_split": {
"dev": {
"questions": 895,
"selection_ce": 0.768348061442913,
"selection_weights": {
"maze/policy": 0.3333333333333333,
"snake/policy": 0.3333333333333333,
"shooting/policy/basic": 0.1556701030927835,
"shooting/policy/predict_position": 0.17766323024054984
},
"missing_selection_cells": [],
"by_task_role": {
"maze/policy": {
"questions": 213,
"excluded_questions": 2,
"objective": "api_policy_distribution",
"ce": 0.8586512628273836,
"kl": 0.051272376984547494,
"tv": 0.12218760235375502,
"brier": 0.03633367460537076,
"target_argmax_agreement": 0.7183098591549296
},
"shooting/policy": {
"questions": 522,
"excluded_questions": 0,
"objective": "mixed_policy_supervision",
"ce": 0.6751665438707917,
"kl": 0.41838145711492203,
"tv": 0.28088536739533787,
"brier": 0.234753920124058,
"target_argmax_agreement": 0.8180076628352491
},
"snake/policy": {
"questions": 157,
"excluded_questions": 1,
"objective": "api_policy_distribution",
"ce": 0.6508974866023702,
"kl": 0.03968017376322264,
"tv": 0.08737390050677335,
"brier": 0.021746081067941795,
"target_argmax_agreement": 0.9044585987261147
}
},
"by_role_macro_task": {
"policy": {
"tasks_present": 3,
"questions": 892,
"ce": 0.7282384311001818,
"kl": 0.16977800262089737,
"tv": 0.16348229008528875,
"brier": 0.09761122526579018,
"target_argmax_agreement": 0.8135920402387645
}
},
"by_selection_pool": {
"maze/policy": {
"questions": 213,
"ce": 0.8586512628273836,
"kl": 0.051272376984547494,
"tv": 0.12218760235375502,
"brier": 0.03633367460537076,
"target_argmax_agreement": 0.7183098591549296
},
"snake/policy": {
"questions": 157,
"ce": 0.6508974866023702,
"kl": 0.03968017376322264,
"tv": 0.08737390050677335,
"brier": 0.021746081067941795,
"target_argmax_agreement": 0.9044585987261147
},
"shooting/policy/predict_position": {
"questions": 138,
"ce": 1.0046922201260415,
"kl": 0.03337471804949119,
"tv": 0.09523188281426646,
"brier": 0.017212267481975678,
"target_argmax_agreement": 0.8333333333333334
},
"shooting/policy/basic": {
"questions": 384,
"ce": 0.5567432539665612,
"kl": 0.5567432539665612,
"tv": 0.34760458841666037,
"brier": 0.3129329515423063,
"target_argmax_agreement": 0.8125
}
},
"by_task_role_target_source": {
"maze/policy/api_policy_distribution": {
"questions": 213,
"ce": 0.8586512628273836,
"kl": 0.051272376984547494,
"tv": 0.12218760235375502,
"brier": 0.03633367460537076,
"target_argmax_agreement": 0.7183098591549296
},
"snake/policy/api_policy_distribution": {
"questions": 157,
"ce": 0.6508974866023702,
"kl": 0.03968017376322264,
"tv": 0.08737390050677335,
"brier": 0.021746081067941795,
"target_argmax_agreement": 0.9044585987261147
},
"shooting/policy/api_policy_distribution": {
"questions": 138,
"ce": 1.0046922201260415,
"kl": 0.03337471804949119,
"tv": 0.09523188281426646,
"brier": 0.017212267481975678,
"target_argmax_agreement": 0.8333333333333334
},
"shooting/policy/expert_action": {
"questions": 384,
"ce": 0.5567432539665612,
"kl": 0.5567432539665612,
"tv": 0.34760458841666037,
"brier": 0.3129329515423063,
"target_argmax_agreement": 0.8125
}
},
"excluded_by_selection_pool": {
"maze/policy": 2,
"snake/policy": 1
},
"notes": "Selection CE uses the explicitly declared population pools. Other task/role metrics are question means within each task. Policy CE/TV/KL measure API or expert target matching, not observed success; target sources are reported separately. Observed Boolean CE/Brier measure event prediction. Binary Brier sums both classes."
},
"calibration": {
"questions": 829,
"selection_ce": 0.7873890376268821,
"selection_weights": {
"maze/policy": 0.3333333333333333,
"snake/policy": 0.3333333333333333,
"shooting/policy/basic": 0.1556701030927835,
"shooting/policy/predict_position": 0.17766323024054984
},
"missing_selection_cells": [],
"by_task_role": {
"maze/policy": {
"questions": 188,
"excluded_questions": 0,
"objective": "api_policy_distribution",
"ce": 0.8557721273005748,
"kl": 0.049116758996135636,
"tv": 0.1178015623817113,
"brier": 0.03488752145795607,
"target_argmax_agreement": 0.7127659574468085
},
"shooting/policy": {
"questions": 503,
"excluded_questions": 0,
"objective": "mixed_policy_supervision",
"ce": 0.7098983267980059,
"kl": 0.45050097122273264,
"tv": 0.29333462310554453,
"brier": 0.2526518407243916,
"target_argmax_agreement": 0.8071570576540755
},
"snake/policy": {
"questions": 137,
"excluded_questions": 1,
"objective": "api_policy_distribution",
"ce": 0.6635669223565426,
"kl": 0.03503055143570352,
"tv": 0.08288389443723143,
"brier": 0.0177465026428542,
"target_argmax_agreement": 0.8467153284671532
}
},
"by_role_macro_task": {
"policy": {
"tasks_present": 3,
"questions": 828,
"ce": 0.7430791254850412,
"kl": 0.17821609388485726,
"tv": 0.1646733599748291,
"brier": 0.10176195494173396,
"target_argmax_agreement": 0.7888794478560124
}
},
"by_selection_pool": {
"maze/policy": {
"questions": 188,
"ce": 0.8557721273005748,
"kl": 0.049116758996135636,
"tv": 0.1178015623817113,
"brier": 0.03488752145795607,
"target_argmax_agreement": 0.7127659574468085
},
"snake/policy": {
"questions": 137,
"ce": 0.6635669223565426,
"kl": 0.03503055143570352,
"tv": 0.08288389443723143,
"brier": 0.0177465026428542,
"target_argmax_agreement": 0.8467153284671532
},
"shooting/policy/predict_position": {
"questions": 126,
"ce": 1.062584408401197,
"kl": 0.027053695271336352,
"tv": 0.0825563326068135,
"brier": 0.01327429250676614,
"target_argmax_agreement": 0.8174603174603174
},
"shooting/policy/basic": {
"questions": 377,
"ce": 0.5920244639810243,
"kl": 0.5920244639810243,
"tv": 0.3637804178080382,
"brier": 0.33265600803320017,
"target_argmax_agreement": 0.8037135278514589
}
},
"by_task_role_target_source": {
"maze/policy/api_policy_distribution": {
"questions": 188,
"ce": 0.8557721273005748,
"kl": 0.049116758996135636,
"tv": 0.1178015623817113,
"brier": 0.03488752145795607,
"target_argmax_agreement": 0.7127659574468085
},
"snake/policy/api_policy_distribution": {
"questions": 137,
"ce": 0.6635669223565426,
"kl": 0.03503055143570352,
"tv": 0.08288389443723143,
"brier": 0.0177465026428542,
"target_argmax_agreement": 0.8467153284671532
},
"shooting/policy/api_policy_distribution": {
"questions": 126,
"ce": 1.062584408401197,
"kl": 0.027053695271336352,
"tv": 0.0825563326068135,
"brier": 0.01327429250676614,
"target_argmax_agreement": 0.8174603174603174
},
"shooting/policy/expert_action": {
"questions": 377,
"ce": 0.5920244639810243,
"kl": 0.5920244639810243,
"tv": 0.3637804178080382,
"brier": 0.33265600803320017,
"target_argmax_agreement": 0.8037135278514589
}
},
"excluded_by_selection_pool": {
"snake/policy": 1
},
"notes": "Selection CE uses the explicitly declared population pools. Other task/role metrics are question means within each task. Policy CE/TV/KL measure API or expert target matching, not observed success; target sources are reported separately. Observed Boolean CE/Brier measure event prediction. Binary Brier sums both classes."
},
"test": {
"questions": 1222,
"selection_ce": 0.8010331335011466,
"selection_weights": {
"maze/policy": 0.3333333333333333,
"snake/policy": 0.3333333333333333,
"shooting/policy/basic": 0.1556701030927835,
"shooting/policy/predict_position": 0.17766323024054984
},
"missing_selection_cells": [],
"by_task_role": {
"maze/policy": {
"questions": 211,
"excluded_questions": 0,
"objective": "api_policy_distribution",
"ce": 0.8259242010118525,
"kl": 0.0512703377492425,
"tv": 0.12187708382354649,
"brier": 0.0353808629987452,
"target_argmax_agreement": 0.7109004739336493
},
"shooting/policy": {
"questions": 895,
"excluded_questions": 0,
"objective": "mixed_policy_supervision",
"ce": 0.6838506166237334,
"kl": 0.5231952973801718,
"tv": 0.3340690186488179,
"brier": 0.2923913738122801,
"target_argmax_agreement": 0.8201117318435754
},
"snake/policy": {
"questions": 116,
"excluded_questions": 0,
"objective": "api_policy_distribution",
"ce": 0.7184771101331556,
"kl": 0.03106495699993995,
"tv": 0.07785939005087449,
"brier": 0.017003446389262544,
"target_argmax_agreement": 0.8879310344827587
}
},
"by_role_macro_task": {
"policy": {
"tasks_present": 3,
"questions": 1222,
"ce": 0.7427506425895806,
"kl": 0.20184353070978475,
"tv": 0.17793516417441294,
"brier": 0.11492522773342928,
"target_argmax_agreement": 0.8063144134199945
}
},
"by_selection_pool": {
"maze/policy": {
"questions": 211,
"ce": 0.8259242010118525,
"kl": 0.0512703377492425,
"tv": 0.12187708382354649,
"brier": 0.0353808629987452,
"target_argmax_agreement": 0.7109004739336493
},
"snake/policy": {
"questions": 116,
"ce": 0.7184771101331556,
"kl": 0.03106495699993995,
"tv": 0.07785939005087449,
"brier": 0.017003446389262544,
"target_argmax_agreement": 0.8879310344827587
},
"shooting/policy/predict_position": {
"questions": 139,
"ce": 1.07489976100609,
"kl": 0.04046443206373267,
"tv": 0.10568520338309749,
"brier": 0.020831237825591608,
"target_argmax_agreement": 0.841726618705036
},
"shooting/policy/basic": {
"questions": 756,
"ce": 0.611951369177771,
"kl": 0.611951369177771,
"tv": 0.3760602227783617,
"brier": 0.34232108135480616,
"target_argmax_agreement": 0.8161375661375662
}
},
"by_task_role_target_source": {
"maze/policy/api_policy_distribution": {
"questions": 211,
"ce": 0.8259242010118525,
"kl": 0.0512703377492425,
"tv": 0.12187708382354649,
"brier": 0.0353808629987452,
"target_argmax_agreement": 0.7109004739336493
},
"snake/policy/api_policy_distribution": {
"questions": 116,
"ce": 0.7184771101331556,
"kl": 0.03106495699993995,
"tv": 0.07785939005087449,
"brier": 0.017003446389262544,
"target_argmax_agreement": 0.8879310344827587
},
"shooting/policy/api_policy_distribution": {
"questions": 139,
"ce": 1.07489976100609,
"kl": 0.04046443206373267,
"tv": 0.10568520338309749,
"brier": 0.020831237825591608,
"target_argmax_agreement": 0.841726618705036
},
"shooting/policy/expert_action": {
"questions": 756,
"ce": 0.611951369177771,
"kl": 0.611951369177771,
"tv": 0.3760602227783617,
"brier": 0.34232108135480616,
"target_argmax_agreement": 0.8161375661375662
}
},
"excluded_by_selection_pool": {},
"notes": "Selection CE uses the explicitly declared population pools. Other task/role metrics are question means within each task. Policy CE/TV/KL measure API or expert target matching, not observed success; target sources are reported separately. Observed Boolean CE/Brier measure event prediction. Binary Brier sums both classes."
},
"ood": {
"questions": 1034,
"selection_ce": 0.8316777569993143,
"selection_weights": {
"maze/policy": 0.3333333333333333,
"snake/policy": 0.3333333333333333,
"shooting/policy/basic": 0.1556701030927835,
"shooting/policy/predict_position": 0.17766323024054984
},
"missing_selection_cells": [],
"by_task_role": {
"maze/policy": {
"questions": 202,
"excluded_questions": 0,
"objective": "api_policy_distribution",
"ce": 0.8944500277474043,
"kl": 0.05899293970612548,
"tv": 0.13182221707821157,
"brier": 0.043880809340294834,
"target_argmax_agreement": 0.6138613861386139
},
"shooting/policy": {
"questions": 688,
"excluded_questions": 1,
"objective": "mixed_policy_supervision",
"ce": 0.711776172345522,
"kl": 0.5716331508554026,
"tv": 0.38142726603535343,
"brier": 0.3145144078339017,
"target_argmax_agreement": 0.8677325581395349
},
"snake/policy": {
"questions": 141,
"excluded_questions": 2,
"objective": "api_policy_distribution",
"ce": 0.7445295586165792,
"kl": 0.06641817329521485,
"tv": 0.11807409310477461,
"brier": 0.0363388908442046,
"target_argmax_agreement": 0.8439716312056738
}
},
"by_role_macro_task": {
"policy": {
"tasks_present": 3,
"questions": 1031,
"ce": 0.7835852529031685,
"kl": 0.23234808795224762,
"tv": 0.21044119207277986,
"brier": 0.13157803600613371,
"target_argmax_agreement": 0.7751885251612741
}
},
"by_selection_pool": {
"maze/policy": {
"questions": 202,
"ce": 0.8944500277474043,
"kl": 0.05899293970612548,
"tv": 0.13182221707821157,
"brier": 0.043880809340294834,
"target_argmax_agreement": 0.6138613861386139
},
"snake/policy": {
"questions": 141,
"ce": 0.7445295586165792,
"kl": 0.06641817329521485,
"tv": 0.11807409310477461,
"brier": 0.0363388908442046,
"target_argmax_agreement": 0.8439716312056738
},
"shooting/policy/predict_position": {
"questions": 99,
"ce": 1.029222704027579,
"kl": 0.05529948397503138,
"tv": 0.13637101541295324,
"brier": 0.03019218594976505,
"target_argmax_agreement": 0.9090909090909091
},
"shooting/policy/basic": {
"questions": 589,
"ce": 0.658419285016959,
"kl": 0.658419285016959,
"tv": 0.4226166867681508,
"brier": 0.36230371168199943,
"target_argmax_agreement": 0.8607809847198642
}
},
"by_task_role_target_source": {
"maze/policy/api_policy_distribution": {
"questions": 202,
"ce": 0.8944500277474043,
"kl": 0.05899293970612548,
"tv": 0.13182221707821157,
"brier": 0.043880809340294834,
"target_argmax_agreement": 0.6138613861386139
},
"snake/policy/api_policy_distribution": {
"questions": 141,
"ce": 0.7445295586165792,
"kl": 0.06641817329521485,
"tv": 0.11807409310477461,
"brier": 0.0363388908442046,
"target_argmax_agreement": 0.8439716312056738
},
"shooting/policy/api_policy_distribution": {
"questions": 99,
"ce": 1.029222704027579,
"kl": 0.05529948397503138,
"tv": 0.13637101541295324,
"brier": 0.03019218594976505,
"target_argmax_agreement": 0.9090909090909091
},
"shooting/policy/expert_action": {
"questions": 589,
"ce": 0.658419285016959,
"kl": 0.658419285016959,
"tv": 0.4226166867681508,
"brier": 0.36230371168199943,
"target_argmax_agreement": 0.8607809847198642
}
},
"excluded_by_selection_pool": {
"shooting/policy/predict_position": 1,
"snake/policy": 2
},
"notes": "Selection CE uses the explicitly declared population pools. Other task/role metrics are question means within each task. Policy CE/TV/KL measure API or expert target matching, not observed success; target sources are reported separately. Observed Boolean CE/Brier measure event prediction. Binary Brier sums both classes."
}
},
"training_seconds": 1716.403577774763,
"max_gpu_allocated_gb": 16.85609728,
"weights_sha256": "38116340795de1c82369b7fe15819d92d79600a7b4dc7a3cd0d4390cb6782639",
"continuation_policy_id": null,
"temperature": 1.0,
"temperature_fitted": false,
"online_compute": {
"forward_calls": 1130,
"question_instances": 7200,
"leaf_paths": 23349,
"padded_tokens": 23040228
},
"td": {
"enabled": false,
"n_step": 3,
"weight": 0.0,
"target_update_every": 25,
"target_refresh_steps": [],
"target_forward_calls": 0,
"target_question_predictions": 0,
"target_leaf_paths": 0,
"target_padded_tokens": 0,
"target_forward_seconds": 0,
"terminal_target_instances": 0,
"bootstrap_target_instances": 0
}
}

Xet Storage Details

Size:
19.3 kB
·
Xet hash:
a7e4a7f3b6bb9724fbdc4f2e126b26dcec0b14010d0de243ec823592a292bb00

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.