{"stage": "restored", "seconds": 147.41660070419312, "init": "tinker://21e32976f18f971463410adff4e364e4/weights/rx19badd842c_init", "info": {"type": "get_info", "model_data": {"arch": null, "model_name": "Qwen/Qwen3.5-2B", "tokenizer_id": null}, "model_id": "daecfa24-b646-4b82-b337-4447b0e95af0:train:0", "optimizer_config": {"type": "adamw"}, "is_lora": true, "lora_rank": 32, "model_name": "Qwen/Qwen3.5-2B"}, "t": 153.31} {"stage": "sampler", "seconds": 3.264300584793091, "sampler_path": "tinker://daecfa24-b646-4b82-b337-4447b0e95af0:train:0/sampler_weights/rx21b_diag_sampler_init", "t": 156.58} {"stage": "sampled", "seconds": 114.1873996257782, "reward_mean": 0.5789062500000001, "n_datums": 256, "t": 270.77} {"stage": "frozen", "batch_sha256": "1e0228e8ad1b59241092728e34d5423d1a4c1d08e53d12bedeef9e354614f5f9", "t": 270.91} {"stage": "sentinel", "tag": "init", "n": 1976, "sha256": "10e156f726b8feacca2c615c4257f50c7cfa06f63a56b676bd9dfbb60806c1da", "vs_sampler_absdp_q90": 0.007642295211553574, "vs_sampler_absdp_max": 0.0621759295463562, "vs_sampler_mean_abs_dlogp": 0.006447661202400923, "t": 333.44} {"n_loss_tokens": 11024, "n_pos_adv": 5680, "n_neg_adv": 5344, "n_zero_adv": 0, "dppo_masked": 0, "dppo_masked_pos": 0, "dppo_masked_neg": 0, "ppo_clipped": 5, "ppo_clipped_pos": 4, "ppo_clipped_neg": 1, "abs_drift_gt_delta": 0, "drift_gt_delta": 0, "drift_lt_neg_delta": 0, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.009247535467147829, "absd_q99": 0.030720190107822423, "absd_max": 0.06226441264152527, "d_min": -0.06226441264152527, "d_max": 0.060592710971832275, "d_q01": -0.02826233744621277, "d_q99": 0.027974664568901075, "d_pos_adv_max": 0.060592710971832275, "d_neg_adv_min": -0.06226441264152527, "ratio_min": 0.6951836347579956, "ratio_max": 1.3449068069458008, "ratio_q99": 1.1022623622417453, "ratio_q01": 0.9058216571807861, "mu_bins": [99, 351, 530, 456, 581, 1003, 1530, 6474], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 11024.0, "n_nonzero_adv_tokens": 11024, "dppo_removed_gradient_tokens": 0, "dppo_masked_nonzero_adv_fraction": 0.0, "policy_drift_abs_p_q99": 0.030720190107822423, "policy_drift_abs_p_max": 0.06226441264152527, "policy_drift_signed_p_min": -0.06226441264152527, "policy_drift_signed_p_max": 0.060592710971832275, "abs_drift_above_delta_tokens": 0, "dppo_masked_token_fraction": 0.0, "grpo_clip_fraction": 0.00045355587808418, "policy_drift_abs_p_q90": 0.009247534908354282, "policy_drift_abs_p_mean": 0.002685021915465825, "mean_ratio": 1.0001580715179443, "nonzero_adv_fraction": 1.0}, "removed_nonzero_grad_pos": 0, "removed_nonzero_grad_neg": 0, "removed_tokens_grad_max_abs": null, "ppo_zero_grad_nonzero_adv": 5, "stage": "update", "update": 1, "epoch": 1, "minibatch": 0, "minibatch_sha256": "e617b4e6e6b67e72bbf107d5121f7253fb574e95b1d7d96e308aa223d0c85954", "n_datums": 32, "intervening_updates_since_sampling": 0, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 0.0, "timing/save_sampler_weights_s": 0.31456787700000177, "timing/save_sampler_weights_calls": 1.0, "timing/idle_wait_s": 113.41697082800005, "timing/idle_wait_calls": 17.0, "timing/publish_weights_s": 1.8178237410000122, "timing/publish_weights_calls": 1.0, "timing/save_checkpoint_s": 5.0993095900000185, "timing/save_checkpoint_calls": 3.0, "timing/persist_checkpoint_s": 9.958550522000024, "timing/persist_checkpoint_calls": 3.0, "timing/prepare_batch_s": 0.007465322999991031, "timing/prepare_batch_calls": 3.0, "timing/forward_only_s": 77.83093168300005, "timing/forward_only_calls": 2.0, "timing/build_outputs_s": 0.0032620620000329836, "timing/build_outputs_calls": 3.0, "timing/forward_backward_s": 25.277396888999988, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.05156856199999993, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 25.328965450999988, "grad_norm:mean": 959.1858520507812}, "t": 380.56} {"stage": "sentinel", "tag": "after_u1", "n": 1976, "sha256": "19f5893d062469b6459b6139ef88258ca479352966a813c076c0756d83e48154", "vs_sampler_absdp_q90": 0.006376773118972778, "vs_sampler_absdp_max": 0.06217598915100098, "vs_sampler_mean_abs_dlogp": 0.005680710077285767, "vs_init_absdp_q90": 0.005858182907104492, "vs_init_absdp_max": 0.06052061915397644, "vs_init_mean_abs_dlogp": 0.005724234972149134, "t": 382.85} {"n_loss_tokens": 10169, "n_pos_adv": 2948, "n_neg_adv": 4430, "n_zero_adv": 2791, "dppo_masked": 0, "dppo_masked_pos": 0, "dppo_masked_neg": 0, "ppo_clipped": 3, "ppo_clipped_pos": 2, "ppo_clipped_neg": 1, "abs_drift_gt_delta": 0, "drift_gt_delta": 0, "drift_lt_neg_delta": 0, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.01059178113937378, "absd_q99": 0.030628552436828597, "absd_max": 0.06178054213523865, "d_min": -0.060083627700805664, "d_max": 0.06178054213523865, "d_q01": -0.028505597114562988, "d_q99": 0.02859009027481079, "d_pos_adv_max": 0.06178054213523865, "d_neg_adv_min": -0.051008909940719604, "ratio_min": 0.7838518023490906, "ratio_max": 1.253517508506775, "ratio_q99": 1.1033843278884887, "ratio_q01": 0.9037534022331238, "mu_bins": [143, 396, 596, 556, 647, 1068, 1561, 5202], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 10169.0, "n_nonzero_adv_tokens": 7378, "dppo_removed_gradient_tokens": 0, "dppo_masked_nonzero_adv_fraction": 0.0, "policy_drift_abs_p_q99": 0.030628552436828597, "policy_drift_abs_p_max": 0.06178054213523865, "policy_drift_signed_p_min": -0.060083627700805664, "policy_drift_signed_p_max": 0.06178054213523865, "abs_drift_above_delta_tokens": 0, "dppo_masked_token_fraction": 0.0, "grpo_clip_fraction": 0.00029501425902251943, "policy_drift_abs_p_q90": 0.01059178076684475, "policy_drift_abs_p_mean": 0.002961999792922598, "mean_ratio": 1.0001107454299927, "nonzero_adv_fraction": 0.725538401022716}, "removed_nonzero_grad_pos": 0, "removed_nonzero_grad_neg": 0, "removed_tokens_grad_max_abs": null, "ppo_zero_grad_nonzero_adv": 3, "stage": "update", "update": 2, "epoch": 1, "minibatch": 1, "minibatch_sha256": "ff8cc7c30bce7afdc37dca59b95a5e6a768412dedec5592cc4cf9289ead2a038", "n_datums": 32, "intervening_updates_since_sampling": 1, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 1.0, "timing/idle_wait_s": 4.368747075999863, "timing/idle_wait_calls": 10.0, "timing/prepare_batch_s": 0.005157371000052535, "timing/prepare_batch_calls": 3.0, "timing/forward_only_s": 2.937602922999986, "timing/forward_only_calls": 2.0, "timing/build_outputs_s": 0.00198544900001707, "timing/build_outputs_calls": 3.0, "timing/forward_backward_s": 5.393086739000012, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.012736782999979823, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 5.4058235219999915, "grad_norm:mean": 479.9132995605469}, "t": 396.07} {"stage": "sentinel", "tag": "after_u2", "n": 1976, "sha256": "99a3fb5d831822fab10f3c4e117d980b5eca7b60a7bf3a2d6eb9cbb78f0cc78a", "vs_sampler_absdp_q90": 0.006704002618789673, "vs_sampler_absdp_max": 0.062101125717163086, "vs_sampler_mean_abs_dlogp": 0.006335264537483454, "vs_init_absdp_q90": 0.007420867681503296, "vs_init_absdp_max": 0.061342209577560425, "vs_init_mean_abs_dlogp": 0.006251731887459755, "t": 397.41} {"n_loss_tokens": 9328, "n_pos_adv": 5208, "n_neg_adv": 4120, "n_zero_adv": 0, "dppo_masked": 0, "dppo_masked_pos": 0, "dppo_masked_neg": 0, "ppo_clipped": 7, "ppo_clipped_pos": 7, "ppo_clipped_neg": 0, "abs_drift_gt_delta": 0, "drift_gt_delta": 0, "drift_lt_neg_delta": 0, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.010938829183578538, "absd_q99": 0.03154855072498321, "absd_max": 0.07395762205123901, "d_min": -0.07395762205123901, "d_max": 0.062215954065322876, "d_q01": -0.02964560627937317, "d_q99": 0.029736233949661253, "d_pos_adv_max": 0.062215954065322876, "d_neg_adv_min": -0.061998069286346436, "ratio_min": 0.783367931842804, "ratio_max": 1.3324962854385376, "ratio_q99": 1.111430444717407, "ratio_q01": 0.9018154627084732, "mu_bins": [93, 309, 434, 468, 564, 790, 1154, 5516], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 9328.0, "n_nonzero_adv_tokens": 9328, "dppo_removed_gradient_tokens": 0, "dppo_masked_nonzero_adv_fraction": 0.0, "policy_drift_abs_p_q99": 0.03154855072498321, "policy_drift_abs_p_max": 0.07395762205123901, "policy_drift_signed_p_min": -0.07395762205123901, "policy_drift_signed_p_max": 0.062215954065322876, "abs_drift_above_delta_tokens": 0, "dppo_masked_token_fraction": 0.0, "grpo_clip_fraction": 0.0007504288164665523, "policy_drift_abs_p_q90": 0.010938819497823715, "policy_drift_abs_p_mean": 0.003040848184677312, "mean_ratio": 1.0002598762512207, "nonzero_adv_fraction": 1.0}, "removed_nonzero_grad_pos": 0, "removed_nonzero_grad_neg": 0, "removed_tokens_grad_max_abs": null, "ppo_zero_grad_nonzero_adv": 7, "stage": "update", "update": 3, "epoch": 1, "minibatch": 2, "minibatch_sha256": "cb0fc21218ffe10d22ce42d964ab03cf31feabb84303acf39370d104a02cc52e", "n_datums": 32, "intervening_updates_since_sampling": 2, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 2.0, "timing/idle_wait_s": 5.724007019999931, "timing/idle_wait_calls": 10.0, "timing/prepare_batch_s": 0.004758284999979878, "timing/prepare_batch_calls": 3.0, "timing/forward_only_s": 0.2510886929999856, "timing/forward_only_calls": 2.0, "timing/build_outputs_s": 0.0019459619999793176, "timing/build_outputs_calls": 3.0, "timing/forward_backward_s": 0.32303479900002685, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.013116680000052838, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.3361514790000797, "grad_norm:mean": 686.292724609375}, "t": 402.54} {"n_loss_tokens": 10318, "n_pos_adv": 3791, "n_neg_adv": 4614, "n_zero_adv": 1913, "dppo_masked": 0, "dppo_masked_pos": 0, "dppo_masked_neg": 0, "ppo_clipped": 7, "ppo_clipped_pos": 4, "ppo_clipped_neg": 3, "abs_drift_gt_delta": 0, "drift_gt_delta": 0, "drift_lt_neg_delta": 0, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.010717272758483937, "absd_q99": 0.03255952984094619, "absd_max": 0.07203680276870728, "d_min": -0.06258267164230347, "d_max": 0.07203680276870728, "d_q01": -0.029504710733890535, "d_q99": 0.029562915563583373, "d_pos_adv_max": 0.07203680276870728, "d_neg_adv_min": -0.06258267164230347, "ratio_min": 0.7322578430175781, "ratio_max": 1.6431528329849243, "ratio_q99": 1.1037465691566468, "ratio_q01": 0.8969535052776336, "mu_bins": [120, 338, 546, 489, 540, 901, 1449, 5935], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 10318.0, "n_nonzero_adv_tokens": 8405, "dppo_removed_gradient_tokens": 0, "dppo_masked_nonzero_adv_fraction": 0.0, "policy_drift_abs_p_q99": 0.03255952984094619, "policy_drift_abs_p_max": 0.07203680276870728, "policy_drift_signed_p_min": -0.06258267164230347, "policy_drift_signed_p_max": 0.07203680276870728, "abs_drift_above_delta_tokens": 0, "dppo_masked_token_fraction": 0.0, "grpo_clip_fraction": 0.0006784260515603799, "policy_drift_abs_p_q90": 0.010717261582612991, "policy_drift_abs_p_mean": 0.0030238982198219884, "mean_ratio": 1.000157117843628, "nonzero_adv_fraction": 0.8145958519092847}, "removed_nonzero_grad_pos": 0, "removed_nonzero_grad_neg": 0, "removed_tokens_grad_max_abs": null, "ppo_zero_grad_nonzero_adv": 7, "stage": "update", "update": 4, "epoch": 1, "minibatch": 3, "minibatch_sha256": "62ac65a25ea157d760d39c62d54ce91f24442dcf7527e8506e8e398e133167a8", "n_datums": 32, "intervening_updates_since_sampling": 3, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 3.0, "timing/idle_wait_s": 4.510585425000045, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.004908081999985825, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.1377093039999977, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.0016823519999888958, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.31578808699998717, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.012830816000018785, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.32861890300000596, "grad_norm:mean": 546.9549560546875}, "t": 407.49} {"stage": "sentinel", "tag": "after_u4", "n": 1976, "sha256": "ed2b5094849eaac7235c07597e38ab011235f4bac9afaf07bdeccab7218582a2", "vs_sampler_absdp_q90": 0.010810106992721558, "vs_sampler_absdp_max": 0.09195864200592041, "vs_sampler_mean_abs_dlogp": 0.008140268735587597, "vs_init_absdp_q90": 0.009358584880828857, "vs_init_absdp_max": 0.09145200252532959, "vs_init_mean_abs_dlogp": 0.007964221760630608, "t": 408.85} {"n_loss_tokens": 9826, "n_pos_adv": 1136, "n_neg_adv": 4979, "n_zero_adv": 3711, "dppo_masked": 0, "dppo_masked_pos": 0, "dppo_masked_neg": 0, "ppo_clipped": 9, "ppo_clipped_pos": 2, "ppo_clipped_neg": 7, "abs_drift_gt_delta": 0, "drift_gt_delta": 0, "drift_lt_neg_delta": 0, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.012813299894332886, "absd_q99": 0.039527490735054016, "absd_max": 0.11470693349838257, "d_min": -0.08419862389564514, "d_max": 0.11470693349838257, "d_q01": -0.030747003853321075, "d_q99": 0.030407339334487915, "d_pos_adv_max": 0.07953274250030518, "d_neg_adv_min": -0.08419862389564514, "ratio_min": 0.6111974120140076, "ratio_max": 1.2590643167495728, "ratio_q99": 1.1157900989055634, "ratio_q01": 0.8940810561180115, "mu_bins": [152, 395, 566, 546, 628, 891, 1282, 5366], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 9826.0, "n_nonzero_adv_tokens": 6115, "dppo_removed_gradient_tokens": 0, "dppo_masked_nonzero_adv_fraction": 0.0, "policy_drift_abs_p_q99": 0.039527490735054016, "policy_drift_abs_p_max": 0.11470693349838257, "policy_drift_signed_p_min": -0.08419862389564514, "policy_drift_signed_p_max": 0.11470693349838257, "abs_drift_above_delta_tokens": 0, "dppo_masked_token_fraction": 0.0, "grpo_clip_fraction": 0.0009159373091797273, "policy_drift_abs_p_q90": 0.012813286855816841, "policy_drift_abs_p_mean": 0.0035554303727879976, "mean_ratio": 0.9998561143875122, "nonzero_adv_fraction": 0.6223285161815592}, "removed_nonzero_grad_pos": 0, "removed_nonzero_grad_neg": 0, "removed_tokens_grad_max_abs": null, "ppo_zero_grad_nonzero_adv": 9, "stage": "update", "update": 5, "epoch": 1, "minibatch": 4, "minibatch_sha256": "74989e50279449cfd0e8cc8777f983e9b62fea38864cb2a7b5cbe0676fe104ca", "n_datums": 32, "intervening_updates_since_sampling": 4, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 4.0, "timing/idle_wait_s": 5.268507881999881, "timing/idle_wait_calls": 10.0, "timing/prepare_batch_s": 0.004600725999978295, "timing/prepare_batch_calls": 3.0, "timing/forward_only_s": 0.24386640400001625, "timing/forward_only_calls": 2.0, "timing/build_outputs_s": 0.002130604000058156, "timing/build_outputs_calls": 3.0, "timing/forward_backward_s": 0.3182997309999678, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.012261691999981394, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.3305614229999492, "grad_norm:mean": 684.9322509765625}, "t": 413.03} {"n_loss_tokens": 12097, "n_pos_adv": 5800, "n_neg_adv": 6297, "n_zero_adv": 0, "dppo_masked": 0, "dppo_masked_pos": 0, "dppo_masked_neg": 0, "ppo_clipped": 32, "ppo_clipped_pos": 19, "ppo_clipped_neg": 13, "abs_drift_gt_delta": 0, "drift_gt_delta": 0, "drift_lt_neg_delta": 0, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.014192834496498103, "absd_q99": 0.04541282653808554, "absd_max": 0.13031619787216187, "d_min": -0.11645084619522095, "d_max": 0.13031619787216187, "d_q01": -0.031279361248016356, "d_q99": 0.031314990520477246, "d_pos_adv_max": 0.13031619787216187, "d_neg_adv_min": -0.1007765531539917, "ratio_min": 0.6150025129318237, "ratio_max": 1.870315432548523, "ratio_q99": 1.1205879878997793, "ratio_q01": 0.89149827003479, "mu_bins": [98, 406, 677, 665, 728, 1103, 1650, 6770], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 12097.0, "n_nonzero_adv_tokens": 12097, "dppo_removed_gradient_tokens": 0, "dppo_masked_nonzero_adv_fraction": 0.0, "policy_drift_abs_p_q99": 0.04541282653808554, "policy_drift_abs_p_max": 0.13031619787216187, "policy_drift_signed_p_min": -0.11645084619522095, "policy_drift_signed_p_max": 0.13031619787216187, "abs_drift_above_delta_tokens": 0, "dppo_masked_token_fraction": 0.0, "grpo_clip_fraction": 0.002645283954699512, "policy_drift_abs_p_q90": 0.01419283077120781, "policy_drift_abs_p_mean": 0.003913545531779178, "mean_ratio": 1.000052809715271, "nonzero_adv_fraction": 1.0}, "removed_nonzero_grad_pos": 0, "removed_nonzero_grad_neg": 0, "removed_tokens_grad_max_abs": null, "ppo_zero_grad_nonzero_adv": 32, "stage": "update", "update": 6, "epoch": 1, "minibatch": 5, "minibatch_sha256": "85db8bf888a3df496253355dc99fe0036afb55d5b364d2471e1d6ba8c0fd4015", "n_datums": 32, "intervening_updates_since_sampling": 5, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 5.0, "timing/idle_wait_s": 4.057837732999985, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.005495559999985744, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.14605220600003577, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.0024336310000308004, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.5140006710000193, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.012224666999998135, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.5262253380000175, "grad_norm:mean": 954.9139404296875}, "t": 418.24} {"n_loss_tokens": 10625, "n_pos_adv": 5197, "n_neg_adv": 4096, "n_zero_adv": 1332, "dppo_masked": 0, "dppo_masked_pos": 0, "dppo_masked_neg": 0, "ppo_clipped": 11, "ppo_clipped_pos": 10, "ppo_clipped_neg": 1, "abs_drift_gt_delta": 0, "drift_gt_delta": 0, "drift_lt_neg_delta": 0, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.016951310634613064, "absd_q99": 0.04314985275268555, "absd_max": 0.1369556486606598, "d_min": -0.1369556486606598, "d_max": 0.11036330461502075, "d_q01": -0.031465837955474855, "d_q99": 0.031462149620056154, "d_pos_adv_max": 0.11036330461502075, "d_neg_adv_min": -0.09337764978408813, "ratio_min": 0.5528027415275574, "ratio_max": 1.4969607591629028, "ratio_q99": 1.1241450262069703, "ratio_q01": 0.889944429397583, "mu_bins": [126, 429, 661, 585, 673, 1093, 1629, 5429], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 10625.0, "n_nonzero_adv_tokens": 9293, "dppo_removed_gradient_tokens": 0, "dppo_masked_nonzero_adv_fraction": 0.0, "policy_drift_abs_p_q99": 0.04314985275268555, "policy_drift_abs_p_max": 0.1369556486606598, "policy_drift_signed_p_min": -0.1369556486606598, "policy_drift_signed_p_max": 0.11036330461502075, "abs_drift_above_delta_tokens": 0, "dppo_masked_token_fraction": 0.0, "grpo_clip_fraction": 0.0010352941176470587, "policy_drift_abs_p_q90": 0.016951290890574455, "policy_drift_abs_p_mean": 0.0043620809807965386, "mean_ratio": 0.9998548030853271, "nonzero_adv_fraction": 0.8746352941176471}, "removed_nonzero_grad_pos": 0, "removed_nonzero_grad_neg": 0, "removed_tokens_grad_max_abs": null, "ppo_zero_grad_nonzero_adv": 11, "stage": "update", "update": 7, "epoch": 1, "minibatch": 6, "minibatch_sha256": "d1f725501890c10c39604b5518d1ed2710dbad764222686711a0c98256a0b423", "n_datums": 32, "intervening_updates_since_sampling": 6, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 6.0, "timing/idle_wait_s": 4.52308745800002, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.005407563999995091, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.13557997499998464, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.0018629220000434543, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.32070837100002336, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.012648882000007688, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.33335725300003105, "grad_norm:mean": 767.5787353515625}, "t": 423.22} {"n_loss_tokens": 11109, "n_pos_adv": 7147, "n_neg_adv": 3962, "n_zero_adv": 0, "dppo_masked": 0, "dppo_masked_pos": 0, "dppo_masked_neg": 0, "ppo_clipped": 27, "ppo_clipped_pos": 20, "ppo_clipped_neg": 7, "abs_drift_gt_delta": 0, "drift_gt_delta": 0, "drift_lt_neg_delta": 0, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.014356115460395816, "absd_q99": 0.049445927143096924, "absd_max": 0.11412730813026428, "d_min": -0.1137809157371521, "d_max": 0.11412730813026428, "d_q01": -0.03673596978187561, "d_q99": 0.03405615329742432, "d_pos_adv_max": 0.11412730813026428, "d_neg_adv_min": -0.11192232370376587, "ratio_min": 0.5669535398483276, "ratio_max": 1.5300076007843018, "ratio_q99": 1.115299391746521, "ratio_q01": 0.8877365922927857, "mu_bins": [89, 260, 474, 472, 573, 906, 1550, 6785], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 11109.0, "n_nonzero_adv_tokens": 11109, "dppo_removed_gradient_tokens": 0, "dppo_masked_nonzero_adv_fraction": 0.0, "policy_drift_abs_p_q99": 0.049445927143096924, "policy_drift_abs_p_max": 0.11412730813026428, "policy_drift_signed_p_min": -0.1137809157371521, "policy_drift_signed_p_max": 0.11412730813026428, "abs_drift_above_delta_tokens": 0, "dppo_masked_token_fraction": 0.0, "grpo_clip_fraction": 0.0024304617877396704, "policy_drift_abs_p_q90": 0.014356114901602268, "policy_drift_abs_p_mean": 0.003962514058656076, "mean_ratio": 0.99970543384552, "nonzero_adv_fraction": 1.0}, "removed_nonzero_grad_pos": 0, "removed_nonzero_grad_neg": 0, "removed_tokens_grad_max_abs": null, "ppo_zero_grad_nonzero_adv": 27, "stage": "update", "update": 8, "epoch": 1, "minibatch": 7, "minibatch_sha256": "ae702595061d76e4a06fc84d9aa4a940b59be9b3cd4288e3b013be8a642799c0", "n_datums": 32, "intervening_updates_since_sampling": 7, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 7.0, "timing/idle_wait_s": 4.922174711000082, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.004880539000055251, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.1327583400000094, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.0019887199999288896, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.3316905259999885, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.01318433799997365, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.34487486399996214, "grad_norm:mean": 968.4353637695312}, "t": 428.6} {"stage": "sentinel", "tag": "after_u8", "n": 1976, "sha256": "c8b000e2a3dd276266660ef42b819530c8220615c96375ce804d37b14305f658", "vs_sampler_absdp_q90": 0.015132233500480652, "vs_sampler_absdp_max": 0.15422680974006653, "vs_sampler_mean_abs_dlogp": 0.01190256793051958, "vs_init_absdp_q90": 0.015591561794281006, "vs_init_absdp_max": 0.1537201702594757, "vs_init_mean_abs_dlogp": 0.012130395509302616, "t": 430.08} {"n_loss_tokens": 11024, "n_pos_adv": 5680, "n_neg_adv": 5344, "n_zero_adv": 0, "dppo_masked": 2, "dppo_masked_pos": 1, "dppo_masked_neg": 1, "ppo_clipped": 167, "ppo_clipped_pos": 60, "ppo_clipped_neg": 107, "abs_drift_gt_delta": 2, "drift_gt_delta": 1, "drift_lt_neg_delta": 1, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.02027362585067749, "absd_q99": 0.06195635467767729, "absd_max": 0.25454404950141907, "d_min": -0.16390733420848846, "d_max": 0.25454404950141907, "d_q01": -0.05505773156881332, "d_q99": 0.046250342130661024, "d_pos_adv_max": 0.25454404950141907, "d_neg_adv_min": -0.16390733420848846, "ratio_min": 0.3880534768104553, "ratio_max": 3.3368654251098633, "ratio_q99": 1.1471128058433537, "ratio_q01": 0.7981848698854447, "mu_bins": [99, 351, 530, 456, 581, 1003, 1530, 6474], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 11024.0, "n_nonzero_adv_tokens": 11024, "dppo_removed_gradient_tokens": 2, "dppo_masked_nonzero_adv_fraction": 0.000181422351233672, "policy_drift_abs_p_q99": 0.06195635467767729, "policy_drift_abs_p_max": 0.25454404950141907, "policy_drift_signed_p_min": -0.16390733420848846, "policy_drift_signed_p_max": 0.25454404950141907, "abs_drift_above_delta_tokens": 2, "dppo_masked_token_fraction": 0.000181422351233672, "grpo_clip_fraction": 0.01514876632801161, "policy_drift_abs_p_q90": 0.02027362585067749, "policy_drift_abs_p_mean": 0.005599213762544386, "mean_ratio": 0.9967764616012573, "nonzero_adv_fraction": 1.0}, "removed_nonzero_grad_pos": 1, "removed_nonzero_grad_neg": 1, "removed_tokens_grad_max_abs": 0.0, "ppo_zero_grad_nonzero_adv": 167, "stage": "update", "update": 9, "epoch": 2, "minibatch": 0, "minibatch_sha256": "e617b4e6e6b67e72bbf107d5121f7253fb574e95b1d7d96e308aa223d0c85954", "n_datums": 32, "intervening_updates_since_sampling": 8, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 8.0, "timing/idle_wait_s": 6.681703029000062, "timing/idle_wait_calls": 10.0, "timing/prepare_batch_s": 0.005356503000086832, "timing/prepare_batch_calls": 3.0, "timing/forward_only_s": 0.2424250200000415, "timing/forward_only_calls": 2.0, "timing/build_outputs_s": 0.00220360200006553, "timing/build_outputs_calls": 3.0, "timing/forward_backward_s": 0.3268847310000069, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.01406330300000036, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.3409480340000073, "grad_norm:mean": 978.2410888671875}, "t": 435.79} {"n_loss_tokens": 10169, "n_pos_adv": 2948, "n_neg_adv": 4430, "n_zero_adv": 2791, "dppo_masked": 3, "dppo_masked_pos": 1, "dppo_masked_neg": 2, "ppo_clipped": 68, "ppo_clipped_pos": 37, "ppo_clipped_neg": 31, "abs_drift_gt_delta": 13, "drift_gt_delta": 1, "drift_lt_neg_delta": 12, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.020218983292579654, "absd_q99": 0.062384855747222884, "absd_max": 0.20896238088607788, "d_min": -0.20896238088607788, "d_max": 0.1507055163383484, "d_q01": -0.0545441460609436, "d_q99": 0.0440905618667602, "d_pos_adv_max": 0.1507055163383484, "d_neg_adv_min": -0.18404129147529602, "ratio_min": 0.42598479986190796, "ratio_max": 1.7750324010849, "ratio_q99": 1.1395601272583007, "ratio_q01": 0.8290457105636597, "mu_bins": [143, 396, 596, 556, 647, 1068, 1561, 5202], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 10169.0, "n_nonzero_adv_tokens": 7378, "dppo_removed_gradient_tokens": 3, "dppo_masked_nonzero_adv_fraction": 0.0004066142586066685, "policy_drift_abs_p_q99": 0.062384855747222884, "policy_drift_abs_p_max": 0.20896238088607788, "policy_drift_signed_p_min": -0.20896238088607788, "policy_drift_signed_p_max": 0.1507055163383484, "abs_drift_above_delta_tokens": 13, "dppo_masked_token_fraction": 0.00029501425902251943, "grpo_clip_fraction": 0.006686989871177107, "policy_drift_abs_p_q90": 0.0202189814299345, "policy_drift_abs_p_mean": 0.005923109473609974, "mean_ratio": 0.9992294311523438, "nonzero_adv_fraction": 0.725538401022716}, "removed_nonzero_grad_pos": 1, "removed_nonzero_grad_neg": 2, "removed_tokens_grad_max_abs": 0.0, "ppo_zero_grad_nonzero_adv": 68, "stage": "update", "update": 10, "epoch": 2, "minibatch": 1, "minibatch_sha256": "ff8cc7c30bce7afdc37dca59b95a5e6a768412dedec5592cc4cf9289ead2a038", "n_datums": 32, "intervening_updates_since_sampling": 9, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 9.0, "timing/idle_wait_s": 4.650788097000088, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.005206760000021404, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.13802018100000168, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.0019691079999688554, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.3234226289999924, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.013835539000012886, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.3372581680000053, "grad_norm:mean": 443.3963623046875}, "t": 440.66} {"n_loss_tokens": 9328, "n_pos_adv": 5208, "n_neg_adv": 4120, "n_zero_adv": 0, "dppo_masked": 9, "dppo_masked_pos": 5, "dppo_masked_neg": 4, "ppo_clipped": 115, "ppo_clipped_pos": 71, "ppo_clipped_neg": 44, "abs_drift_gt_delta": 20, "drift_gt_delta": 7, "drift_lt_neg_delta": 13, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.02088937461376191, "absd_q99": 0.076246239542961, "absd_max": 0.3603789806365967, "d_min": -0.3603789806365967, "d_max": 0.35609227418899536, "d_q01": -0.05625966817140579, "d_q99": 0.05502174228429762, "d_pos_adv_max": 0.35609227418899536, "d_neg_adv_min": -0.3450714945793152, "ratio_min": 0.4527239501476288, "ratio_max": 3.2908339500427246, "ratio_q99": 1.1869185829162596, "ratio_q01": 0.8330286794900894, "mu_bins": [93, 309, 434, 468, 564, 790, 1154, 5516], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 9328.0, "n_nonzero_adv_tokens": 9328, "dppo_removed_gradient_tokens": 9, "dppo_masked_nonzero_adv_fraction": 0.0009648370497427101, "policy_drift_abs_p_q99": 0.076246239542961, "policy_drift_abs_p_max": 0.3603789806365967, "policy_drift_signed_p_min": -0.3603789806365967, "policy_drift_signed_p_max": 0.35609228909015656, "abs_drift_above_delta_tokens": 20, "dppo_masked_token_fraction": 0.0009648370497427101, "grpo_clip_fraction": 0.012328473413379074, "policy_drift_abs_p_q90": 0.020889373496174812, "policy_drift_abs_p_mean": 0.006231911445471328, "mean_ratio": 0.9996479153633118, "nonzero_adv_fraction": 1.0}, "removed_nonzero_grad_pos": 5, "removed_nonzero_grad_neg": 4, "removed_tokens_grad_max_abs": 0.0, "ppo_zero_grad_nonzero_adv": 115, "stage": "update", "update": 11, "epoch": 2, "minibatch": 2, "minibatch_sha256": "cb0fc21218ffe10d22ce42d964ab03cf31feabb84303acf39370d104a02cc52e", "n_datums": 32, "intervening_updates_since_sampling": 10, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 10.0, "timing/idle_wait_s": 3.418704694999974, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.005170007000003807, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.13357315799999014, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.0016601519999994707, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.3167008739999915, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.013238890000025094, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.32993976400001657, "grad_norm:mean": 682.4713745117188}, "t": 444.53} {"n_loss_tokens": 10318, "n_pos_adv": 3791, "n_neg_adv": 4614, "n_zero_adv": 1913, "dppo_masked": 7, "dppo_masked_pos": 2, "dppo_masked_neg": 5, "ppo_clipped": 93, "ppo_clipped_pos": 51, "ppo_clipped_neg": 42, "abs_drift_gt_delta": 17, "drift_gt_delta": 3, "drift_lt_neg_delta": 14, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.021940040588378933, "absd_q99": 0.08591193854808804, "absd_max": 0.2582799196243286, "d_min": -0.2582799196243286, "d_max": 0.21189898252487183, "d_q01": -0.06027993410825729, "d_q99": 0.05667384415864944, "d_pos_adv_max": 0.17151719331741333, "d_neg_adv_min": -0.242240309715271, "ratio_min": 0.34822171926498413, "ratio_max": 3.538233757019043, "ratio_q99": 1.1767856681346893, "ratio_q01": 0.8254920834302902, "mu_bins": [120, 338, 546, 489, 540, 901, 1449, 5935], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 10318.0, "n_nonzero_adv_tokens": 8405, "dppo_removed_gradient_tokens": 7, "dppo_masked_nonzero_adv_fraction": 0.0008328375966686496, "policy_drift_abs_p_q99": 0.08591193854808804, "policy_drift_abs_p_max": 0.2582799196243286, "policy_drift_signed_p_min": -0.2582799196243286, "policy_drift_signed_p_max": 0.21189898252487183, "abs_drift_above_delta_tokens": 17, "dppo_masked_token_fraction": 0.0006784260515603799, "grpo_clip_fraction": 0.009013374685016477, "policy_drift_abs_p_q90": 0.02194003388285637, "policy_drift_abs_p_mean": 0.006682662220184212, "mean_ratio": 0.9997818470001221, "nonzero_adv_fraction": 0.8145958519092847}, "removed_nonzero_grad_pos": 2, "removed_nonzero_grad_neg": 5, "removed_tokens_grad_max_abs": 0.0, "ppo_zero_grad_nonzero_adv": 93, "stage": "update", "update": 12, "epoch": 2, "minibatch": 3, "minibatch_sha256": "62ac65a25ea157d760d39c62d54ce91f24442dcf7527e8506e8e398e133167a8", "n_datums": 32, "intervening_updates_since_sampling": 11, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 11.0, "timing/idle_wait_s": 3.1966524669999785, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.004634505999945304, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.13577591500001063, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.0017678180000189059, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.3233261960000391, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.013047928999981195, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.3363741250000203, "grad_norm:mean": 547.5277709960938}, "t": 448.39} {"n_loss_tokens": 9826, "n_pos_adv": 1136, "n_neg_adv": 4979, "n_zero_adv": 3711, "dppo_masked": 8, "dppo_masked_pos": 4, "dppo_masked_neg": 4, "ppo_clipped": 122, "ppo_clipped_pos": 41, "ppo_clipped_neg": 81, "abs_drift_gt_delta": 31, "drift_gt_delta": 18, "drift_lt_neg_delta": 13, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.024678528308868408, "absd_q99": 0.08629753440618515, "absd_max": 0.30098220705986023, "d_min": -0.22848473489284515, "d_max": 0.30098220705986023, "d_q01": -0.06204549968242645, "d_q99": 0.06023888289928436, "d_pos_adv_max": 0.30098220705986023, "d_neg_adv_min": -0.22848473489284515, "ratio_min": 0.1927693784236908, "ratio_max": 2.348411798477173, "ratio_q99": 1.20603346824646, "ratio_q01": 0.78770412504673, "mu_bins": [152, 395, 566, 546, 628, 891, 1282, 5366], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 9826.0, "n_nonzero_adv_tokens": 6115, "dppo_removed_gradient_tokens": 8, "dppo_masked_nonzero_adv_fraction": 0.0013082583810302534, "policy_drift_abs_p_q99": 0.08629753440618515, "policy_drift_abs_p_max": 0.30098220705986023, "policy_drift_signed_p_min": -0.22848473489284515, "policy_drift_signed_p_max": 0.30098220705986023, "abs_drift_above_delta_tokens": 31, "dppo_masked_token_fraction": 0.0008141664970486464, "grpo_clip_fraction": 0.012416039079991859, "policy_drift_abs_p_q90": 0.024678517132997513, "policy_drift_abs_p_mean": 0.0074309974250664395, "mean_ratio": 0.998245358467102, "nonzero_adv_fraction": 0.6223285161815592}, "removed_nonzero_grad_pos": 4, "removed_nonzero_grad_neg": 4, "removed_tokens_grad_max_abs": 0.0, "ppo_zero_grad_nonzero_adv": 122, "stage": "update", "update": 13, "epoch": 2, "minibatch": 4, "minibatch_sha256": "74989e50279449cfd0e8cc8777f983e9b62fea38864cb2a7b5cbe0676fe104ca", "n_datums": 32, "intervening_updates_since_sampling": 12, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 12.0, "timing/idle_wait_s": 3.393326320000085, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.0046679449999942335, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.13438858699998946, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.001764240999989397, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.3173125929999969, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.012573879999990822, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.32988647299998775, "grad_norm:mean": 700.5138549804688}, "t": 452.07} {"n_loss_tokens": 12097, "n_pos_adv": 5800, "n_neg_adv": 6297, "n_zero_adv": 0, "dppo_masked": 26, "dppo_masked_pos": 15, "dppo_masked_neg": 11, "ppo_clipped": 230, "ppo_clipped_pos": 128, "ppo_clipped_neg": 102, "abs_drift_gt_delta": 41, "drift_gt_delta": 21, "drift_lt_neg_delta": 20, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.026802468299865715, "absd_q99": 0.09108320474624618, "absd_max": 0.41509950160980225, "d_min": -0.41509950160980225, "d_max": 0.39258474111557007, "d_q01": -0.06893085360527039, "d_q99": 0.06367182493209785, "d_pos_adv_max": 0.39258474111557007, "d_neg_adv_min": -0.3994252383708954, "ratio_min": 0.13229867815971375, "ratio_max": 7.288299083709717, "ratio_q99": 1.2579090929031362, "ratio_q01": 0.7910798192024231, "mu_bins": [98, 406, 677, 665, 728, 1103, 1650, 6770], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 12097.0, "n_nonzero_adv_tokens": 12097, "dppo_removed_gradient_tokens": 26, "dppo_masked_nonzero_adv_fraction": 0.0021492932131933536, "policy_drift_abs_p_q99": 0.09108320474624618, "policy_drift_abs_p_max": 0.41509950160980225, "policy_drift_signed_p_min": -0.41509950160980225, "policy_drift_signed_p_max": 0.39258474111557007, "abs_drift_above_delta_tokens": 41, "dppo_masked_token_fraction": 0.0021492932131933536, "grpo_clip_fraction": 0.019012978424402745, "policy_drift_abs_p_q90": 0.026802463456988335, "policy_drift_abs_p_mean": 0.008372399947754095, "mean_ratio": 1.0001839399337769, "nonzero_adv_fraction": 1.0}, "removed_nonzero_grad_pos": 15, "removed_nonzero_grad_neg": 11, "removed_tokens_grad_max_abs": 0.0, "ppo_zero_grad_nonzero_adv": 230, "stage": "update", "update": 14, "epoch": 2, "minibatch": 5, "minibatch_sha256": "85db8bf888a3df496253355dc99fe0036afb55d5b364d2471e1d6ba8c0fd4015", "n_datums": 32, "intervening_updates_since_sampling": 13, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 13.0, "timing/idle_wait_s": 3.7102522229999977, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.005252467999980581, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.13571139100002938, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.002388642000028085, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.3339291539999749, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.01350042299998222, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.34742957699995713, "grad_norm:mean": 966.0278930664062}, "t": 456.26} {"n_loss_tokens": 10625, "n_pos_adv": 5197, "n_neg_adv": 4096, "n_zero_adv": 1332, "dppo_masked": 16, "dppo_masked_pos": 7, "dppo_masked_neg": 9, "ppo_clipped": 152, "ppo_clipped_pos": 81, "ppo_clipped_neg": 71, "abs_drift_gt_delta": 34, "drift_gt_delta": 9, "drift_lt_neg_delta": 25, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.0267482876777649, "absd_q99": 0.09148484468460083, "absd_max": 0.2911640405654907, "d_min": -0.2911640405654907, "d_max": 0.24314269423484802, "d_q01": -0.07036675572395325, "d_q99": 0.06078783035278328, "d_pos_adv_max": 0.24314269423484802, "d_neg_adv_min": -0.2779943346977234, "ratio_min": 0.22648665308952332, "ratio_max": 2.558509588241577, "ratio_q99": 1.2359069013595583, "ratio_q01": 0.7830280685424805, "mu_bins": [126, 429, 661, 585, 673, 1093, 1629, 5429], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 10625.0, "n_nonzero_adv_tokens": 9293, "dppo_removed_gradient_tokens": 16, "dppo_masked_nonzero_adv_fraction": 0.0017217260303454213, "policy_drift_abs_p_q99": 0.09148484468460083, "policy_drift_abs_p_max": 0.2911640405654907, "policy_drift_signed_p_min": -0.2911640405654907, "policy_drift_signed_p_max": 0.24314269423484802, "abs_drift_above_delta_tokens": 34, "dppo_masked_token_fraction": 0.0015058823529411766, "grpo_clip_fraction": 0.014305882352941177, "policy_drift_abs_p_q90": 0.026748282834887505, "policy_drift_abs_p_mean": 0.008546106274770529, "mean_ratio": 0.9987276792526245, "nonzero_adv_fraction": 0.8746352941176471}, "removed_nonzero_grad_pos": 7, "removed_nonzero_grad_neg": 9, "removed_tokens_grad_max_abs": 0.0, "ppo_zero_grad_nonzero_adv": 152, "stage": "update", "update": 15, "epoch": 2, "minibatch": 6, "minibatch_sha256": "d1f725501890c10c39604b5518d1ed2710dbad764222686711a0c98256a0b423", "n_datums": 32, "intervening_updates_since_sampling": 14, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 14.0, "timing/idle_wait_s": 3.1480931730000634, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.005059779000077924, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.1407873279999876, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.0019973860000277455, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.33606349100000443, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.013373459000035837, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.34943695000004027, "grad_norm:mean": 764.8272094726562}, "t": 460.35} {"n_loss_tokens": 11109, "n_pos_adv": 7147, "n_neg_adv": 3962, "n_zero_adv": 0, "dppo_masked": 25, "dppo_masked_pos": 15, "dppo_masked_neg": 10, "ppo_clipped": 210, "ppo_clipped_pos": 106, "ppo_clipped_neg": 104, "abs_drift_gt_delta": 66, "drift_gt_delta": 19, "drift_lt_neg_delta": 47, "ratio_cap_hits": 0, "ratio_cap_hits_nonzero_adv": 0, "absd_q90": 0.028433847427368167, "absd_q99": 0.11075854063034057, "absd_max": 0.41873541474342346, "d_min": -0.41873541474342346, "d_max": 0.25539734959602356, "d_q01": -0.08800760746002197, "d_q99": 0.07505983591079715, "d_pos_adv_max": 0.25539734959602356, "d_neg_adv_min": -0.4168769419193268, "ratio_min": 0.3299494981765747, "ratio_max": 3.423377275466919, "ratio_q99": 1.22781316280365, "ratio_q01": 0.7354044842720032, "mu_bins": [89, 260, 474, 472, 573, 906, 1550, 6785], "dtype": "torch.float32", "patched_diag": {"n_loss_tokens": 11109.0, "n_nonzero_adv_tokens": 11109, "dppo_removed_gradient_tokens": 25, "dppo_masked_nonzero_adv_fraction": 0.0022504275812404357, "policy_drift_abs_p_q99": 0.11075854063034057, "policy_drift_abs_p_max": 0.41873541474342346, "policy_drift_signed_p_min": -0.41873541474342346, "policy_drift_signed_p_max": 0.25539734959602356, "abs_drift_above_delta_tokens": 66, "dppo_masked_token_fraction": 0.0022504275812404357, "grpo_clip_fraction": 0.01890359168241966, "policy_drift_abs_p_q90": 0.028433846309781075, "policy_drift_abs_p_mean": 0.00929815229572715, "mean_ratio": 0.9973668456077576, "nonzero_adv_fraction": 1.0}, "removed_nonzero_grad_pos": 15, "removed_nonzero_grad_neg": 10, "removed_tokens_grad_max_abs": 0.0, "ppo_zero_grad_nonzero_adv": 210, "stage": "update", "update": 16, "epoch": 2, "minibatch": 7, "minibatch_sha256": "ae702595061d76e4a06fc84d9aa4a940b59be9b3cd4288e3b013be8a642799c0", "n_datums": 32, "intervening_updates_since_sampling": 15, "optim_metrics": {"update_successful:mean": 1.0, "timing/optimizer_step": 15.0, "timing/idle_wait_s": 4.168564484999877, "timing/idle_wait_calls": 7.0, "timing/prepare_batch_s": 0.005365282000013849, "timing/prepare_batch_calls": 2.0, "timing/forward_only_s": 0.13685142799999994, "timing/forward_only_calls": 1.0, "timing/build_outputs_s": 0.0018902119999779643, "timing/build_outputs_calls": 2.0, "timing/forward_backward_s": 0.33104679600000964, "timing/forward_backward_calls": 1.0, "timing/optim_step_s": 0.013079152999978305, "timing/optim_step_calls": 1.0, "timing/trainer_step_s": 0.34412594899998794, "grad_norm:mean": 975.4917602539062}, "t": 464.81} {"stage": "sentinel", "tag": "after_u16", "n": 1976, "sha256": "53ddab1bf5d7be91b7476fd643edb0687a0c3a763ba36b168d5acddd75fae561", "vs_sampler_absdp_q90": 0.03070691227912903, "vs_sampler_absdp_max": 0.3288673162460327, "vs_sampler_mean_abs_dlogp": 0.027858657762408257, "vs_init_absdp_q90": 0.03011302649974823, "vs_init_absdp_max": 0.3283606767654419, "vs_init_mean_abs_dlogp": 0.027917437255382538, "t": 465.91} {"stage": "saved_state", "path": "tinker://daecfa24-b646-4b82-b337-4447b0e95af0:train:0/weights/rx21b_diag_final", "seconds": 4.876049757003784, "t": 470.79} {"stage": "done", "updates": 16, "wall": 470.7910211086273, "t": 470.79}