diff options
Diffstat (limited to 'results/kp_innovation_stability_gate.json')
| -rw-r--r-- | results/kp_innovation_stability_gate.json | 123 |
1 files changed, 123 insertions, 0 deletions
diff --git a/results/kp_innovation_stability_gate.json b/results/kp_innovation_stability_gate.json new file mode 100644 index 0000000..b68a6c0 --- /dev/null +++ b/results/kp_innovation_stability_gate.json @@ -0,0 +1,123 @@ +{ + "checks": { + "0p001": { + "final_32_mean_loss_at_most_2p5": false, + "forward_and_feedback_momentum_max_at_most_50": false, + "forward_and_feedback_weight_max_at_most_10": false, + "maximum_batch_loss_at_most_10": true, + "maximum_used_instruction_rms_ratio_at_most_2": false, + "no_held_out_evaluations": true, + "one_sided_closed_form_certificate": true, + "predictor_frozen_during_task": true, + "record_trajectory_and_state_finite": false, + "task_loader_state_restored": true, + "traffic_ratio_calibrated": true + }, + "0p003": { + "final_32_mean_loss_at_most_2p5": false, + "forward_and_feedback_momentum_max_at_most_50": false, + "forward_and_feedback_weight_max_at_most_10": false, + "maximum_batch_loss_at_most_10": true, + "maximum_used_instruction_rms_ratio_at_most_2": false, + "no_held_out_evaluations": true, + "one_sided_closed_form_certificate": true, + "predictor_frozen_during_task": true, + "record_trajectory_and_state_finite": false, + "task_loader_state_restored": true, + "traffic_ratio_calibrated": true + }, + "0p01": { + "final_32_mean_loss_at_most_2p5": true, + "forward_and_feedback_momentum_max_at_most_50": false, + "forward_and_feedback_weight_max_at_most_10": false, + "maximum_batch_loss_at_most_10": false, + "maximum_used_instruction_rms_ratio_at_most_2": false, + "no_held_out_evaluations": true, + "one_sided_closed_form_certificate": true, + "predictor_frozen_during_task": true, + "record_trajectory_and_state_finite": true, + "task_loader_state_restored": true, + "traffic_ratio_calibrated": true + }, + "0p03": { + "final_32_mean_loss_at_most_2p5": true, + "forward_and_feedback_momentum_max_at_most_50": false, + "forward_and_feedback_weight_max_at_most_10": false, + "maximum_batch_loss_at_most_10": false, + "maximum_used_instruction_rms_ratio_at_most_2": false, + "no_held_out_evaluations": true, + "one_sided_closed_form_certificate": true, + "predictor_frozen_during_task": true, + "record_trajectory_and_state_finite": true, + "task_loader_state_restored": true, + "traffic_ratio_calibrated": true + } + }, + "metrics": { + "0p001": { + "eligible": false, + "final_32_mean_loss": 2.5811559855937958, + "margin": 0.001, + "max_positive_residual_soma_slope": 0.0, + "maximum_batch_loss": 3.3464300632476807, + "maximum_feedback_momentum": 6.509369970445517e+16, + "maximum_feedback_weight": 8.15819417148457e+16, + "maximum_forward_momentum": 6.509369970445517e+16, + "maximum_forward_weight": 8.15819417148457e+16, + "maximum_initial_traffic_ratio_error": 4.76837158203125e-07, + "maximum_used_instruction_rms_ratio": 2413069693172539.5, + "min_residual_soma_slope": -0.0011125205783173442 + }, + "0p003": { + "eligible": false, + "final_32_mean_loss": 4.165377259254456, + "margin": 0.003, + "max_positive_residual_soma_slope": 0.0, + "maximum_batch_loss": 9.148499488830566, + "maximum_feedback_momentum": 1.2453761044919091e+17, + "maximum_feedback_weight": 1.4783445795550003e+17, + "maximum_forward_momentum": 1.2453761044919091e+17, + "maximum_forward_weight": 1.4783445795550003e+17, + "maximum_initial_traffic_ratio_error": 4.76837158203125e-07, + "maximum_used_instruction_rms_ratio": 37248560711075.516, + "min_residual_soma_slope": -0.003337553469464183 + }, + "0p01": { + "eligible": false, + "final_32_mean_loss": 2.3050725758075714, + "margin": 0.01, + "max_positive_residual_soma_slope": 0.0, + "maximum_batch_loss": 139979.8125, + "maximum_feedback_momentum": 8878180.0, + "maximum_feedback_weight": 8827973.0, + "maximum_forward_momentum": 8878180.0, + "maximum_forward_weight": 8827973.0, + "maximum_initial_traffic_ratio_error": 4.76837158203125e-07, + "maximum_used_instruction_rms_ratio": 538.3622583067088, + "min_residual_soma_slope": -0.011125171557068825 + }, + "0p03": { + "eligible": false, + "final_32_mean_loss": 2.317491501569748, + "margin": 0.03, + "max_positive_residual_soma_slope": 0.0, + "maximum_batch_loss": 4965542.0, + "maximum_feedback_momentum": 119389248.0, + "maximum_feedback_weight": 118714088.0, + "maximum_forward_momentum": 119389248.0, + "maximum_forward_weight": 118714088.0, + "maximum_initial_traffic_ratio_error": 4.76837158203125e-07, + "maximum_used_instruction_rms_ratio": 6.592793428967176, + "min_residual_soma_slope": -0.03337550908327103 + } + }, + "protocol": "kp_stability_margin_training_prefix_v1", + "review_score_after": 5, + "review_score_before": 5, + "score_change_rule": "training-only stability selection cannot change the paper score", + "selected": null, + "source_commit": "0ae690e7b3144da2c5b3384cd7169e86529fd26f", + "status": "failed_no_eligible_margin", + "test_evaluations": 0, + "validation_evaluations": 0 +} |
