diff --git "a/trainer_state.json" "b/trainer_state.json"
--- "a/trainer_state.json"
+++ "b/trainer_state.json"
@@ -2,9 +2,9 @@
   "best_global_step": null,
   "best_metric": null,
   "best_model_checkpoint": null,
-  "epoch": 0.00075,
+  "epoch": 0.003,
   "eval_steps": 500,
-  "global_step": 75,
+  "global_step": 300,
   "is_hyper_param_search": false,
   "is_local_process_zero": true,
   "is_world_process_zero": true,
@@ -16,41 +16,41 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 3.0,
-      "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.75,
-      "completions/mean_terminated_length": 2.75,
+      "completions/max_length": 8.0,
+      "completions/max_terminated_length": 8.0,
+      "completions/mean_length": 2.5625,
+      "completions/mean_terminated_length": 2.5625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.8766089752316475,
+      "entropy": 0.8400783389806747,
       "epoch": 1e-05,
-      "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.00977164227515459,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0,
       "kl": 0.0,
       "learning_rate": 0.0,
       "loss": 0.0,
-      "num_tokens": 41130.0,
-      "reward": 2.075000047683716,
-      "reward_std": 1.008032202720642,
+      "num_tokens": 51438.0,
+      "reward": 1.5750000476837158,
+      "reward_std": 1.4312187433242798,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.75,
-      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 1.4312187433242798,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.8125,
-      "rewards/probe_shaping_dominance/std": 0.3965577781200409,
-      "rewards/probe_terminal_raw/mean": 0.1875,
-      "rewards/probe_terminal_raw/std": 0.3965577781200409,
-      "rewards/rollout_reward_func/mean": -0.625,
-      "rewards/rollout_reward_func/std": 0.7931155562400818,
-      "sampling/importance_sampling_ratio/max": 1.4458240270614624,
-      "sampling/importance_sampling_ratio/mean": 0.8427327871322632,
-      "sampling/importance_sampling_ratio/min": 0.452865332365036,
-      "sampling/sampling_logp_difference/max": 0.3974947929382324,
-      "sampling/sampling_logp_difference/mean": 0.08604969829320908,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.0456675291061401,
+      "sampling/importance_sampling_ratio/mean": 0.8575431704521179,
+      "sampling/importance_sampling_ratio/min": 0.5433069467544556,
+      "sampling/sampling_logp_difference/max": 0.4106518030166626,
+      "sampling/sampling_logp_difference/mean": 0.07245150208473206,
       "step": 1,
-      "step_time": 10.943891547999556
+      "step_time": 11.059172731000217
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -58,14 +58,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.8766089752316475,
+      "entropy": 0.8400783389806747,
       "epoch": 2e-05,
-      "grad_norm": 0.009746627882122993,
+      "grad_norm": 0.0,
       "kl": 0.0,
       "learning_rate": 2.2857142857142855e-07,
       "loss": 0.0,
       "step": 2,
-      "step_time": 5.863016629001322
+      "step_time": 6.07740226399801
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -76,39 +76,39 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.84375,
-      "completions/mean_terminated_length": 2.84375,
+      "completions/mean_length": 2.8125,
+      "completions/mean_terminated_length": 2.8125,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.6090607978403568,
+      "entropy": 0.7734263241291046,
       "epoch": 3e-05,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 1.7606014807824977e-05,
-      "kl": 0.002607578739684868,
+      "grad_norm": 0.0,
+      "kl": 0.0,
       "learning_rate": 4.571428571428571e-07,
       "loss": 0.0,
-      "num_tokens": 83246.0,
-      "reward": 2.293750047683716,
-      "reward_std": 1.035167932510376,
+      "num_tokens": 99159.0,
+      "reward": 1.7625000476837158,
+      "reward_std": 0.3965577781200409,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.84375,
-      "rewards/probe_completion_length/std": 0.3689020276069641,
+      "rewards/probe_completion_length/mean": 1.8125,
+      "rewards/probe_completion_length/std": 0.3965577781200409,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.75,
-      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
-      "rewards/probe_terminal_raw/mean": 0.25,
-      "rewards/probe_terminal_raw/std": 0.4399413466453552,
-      "rewards/rollout_reward_func/mean": -0.5,
-      "rewards/rollout_reward_func/std": 0.8798826932907104,
-      "sampling/importance_sampling_ratio/max": 1.640177845954895,
-      "sampling/importance_sampling_ratio/mean": 0.9333716630935669,
-      "sampling/importance_sampling_ratio/min": 0.47254422307014465,
-      "sampling/sampling_logp_difference/max": 0.5935717821121216,
-      "sampling/sampling_logp_difference/mean": 0.0694335401058197,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.0850459337234497,
+      "sampling/importance_sampling_ratio/mean": 0.8818035125732422,
+      "sampling/importance_sampling_ratio/min": 0.6340073347091675,
+      "sampling/sampling_logp_difference/max": 0.315462589263916,
+      "sampling/sampling_logp_difference/mean": 0.06536750495433807,
       "step": 3,
-      "step_time": 10.9592392880013
+      "step_time": 11.278225346995896
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -116,14 +116,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.6121835336089134,
+      "entropy": 0.7734263241291046,
       "epoch": 4e-05,
-      "grad_norm": 7.779162842780352e-05,
-      "kl": 0.005520817290744162,
+      "grad_norm": 0.0,
+      "kl": 0.0,
       "learning_rate": 6.857142857142857e-07,
       "loss": 0.0,
       "step": 4,
-      "step_time": 5.935810445001152
+      "step_time": 5.904026553995209
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -134,24 +134,24 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.46875,
-      "completions/mean_terminated_length": 2.46875,
+      "completions/mean_length": 2.71875,
+      "completions/mean_terminated_length": 2.71875,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.8703878000378609,
+      "entropy": 0.7801894694566727,
       "epoch": 5e-05,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 9.640563075663522e-05,
-      "kl": 0.006076000264101822,
+      "grad_norm": 0.0,
+      "kl": 0.0,
       "learning_rate": 9.142857142857142e-07,
       "loss": 0.0,
-      "num_tokens": 131563.0,
-      "reward": 1.4187500476837158,
-      "reward_std": 0.507007360458374,
+      "num_tokens": 148168.0,
+      "reward": 1.6687500476837158,
+      "reward_std": 0.45680341124534607,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.46875,
-      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_completion_length/mean": 1.71875,
+      "rewards/probe_completion_length/std": 0.45680341124534607,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
       "rewards/probe_shaping_dominance/mean": 1.0,
@@ -160,13 +160,13 @@
       "rewards/probe_terminal_raw/std": 0.0,
       "rewards/rollout_reward_func/mean": -1.0,
       "rewards/rollout_reward_func/std": 0.0,
-      "sampling/importance_sampling_ratio/max": 1.9647347927093506,
-      "sampling/importance_sampling_ratio/mean": 0.9198578596115112,
-      "sampling/importance_sampling_ratio/min": 0.5797430872917175,
-      "sampling/sampling_logp_difference/max": 0.7410016059875488,
-      "sampling/sampling_logp_difference/mean": 0.0826396569609642,
+      "sampling/importance_sampling_ratio/max": 1.526421070098877,
+      "sampling/importance_sampling_ratio/mean": 0.9378703832626343,
+      "sampling/importance_sampling_ratio/min": 0.6182330846786499,
+      "sampling/sampling_logp_difference/max": 0.8640813827514648,
+      "sampling/sampling_logp_difference/mean": 0.09040169417858124,
       "step": 5,
-      "step_time": 10.165990399998918
+      "step_time": 10.348131929999
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -174,14 +174,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.8728083595633507,
+      "entropy": 0.7801894694566727,
       "epoch": 6e-05,
-      "grad_norm": 7.622271368745714e-05,
-      "kl": 0.009481642691298475,
+      "grad_norm": 0.0,
+      "kl": 0.0,
       "learning_rate": 1.1428571428571428e-06,
       "loss": 0.0,
       "step": 6,
-      "step_time": 6.0723176140008945
+      "step_time": 6.946818038999481
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -190,56 +190,56 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 3.0,
-      "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.46875,
-      "completions/mean_terminated_length": 2.46875,
+      "completions/max_length": 7.0,
+      "completions/max_terminated_length": 7.0,
+      "completions/mean_length": 3.03125,
+      "completions/mean_terminated_length": 3.03125,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 1.0045497044920921,
+      "entropy": 0.6854325383901596,
       "epoch": 7e-05,
-      "frac_reward_zero_std": 1.0,
-      "grad_norm": 7.399460446322337e-05,
-      "kl": 0.006289539747740491,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.006476345006376505,
+      "kl": 0.0,
       "learning_rate": 1.3714285714285715e-06,
       "loss": 0.0,
-      "num_tokens": 180273.0,
-      "reward": 1.4187500476837158,
-      "reward_std": 0.507007360458374,
+      "num_tokens": 195423.0,
+      "reward": 2.1062498092651367,
+      "reward_std": 1.167002558708191,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.46875,
-      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_completion_length/mean": 2.09375,
+      "rewards/probe_completion_length/std": 1.1175830364227295,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 1.0,
-      "rewards/probe_shaping_dominance/std": 0.0,
-      "rewards/probe_terminal_raw/mean": 0.0,
-      "rewards/probe_terminal_raw/std": 0.0,
-      "rewards/rollout_reward_func/mean": -1.0,
-      "rewards/rollout_reward_func/std": 0.0,
-      "sampling/importance_sampling_ratio/max": 1.3054357767105103,
-      "sampling/importance_sampling_ratio/mean": 0.8570933938026428,
-      "sampling/importance_sampling_ratio/min": 0.5681248903274536,
-      "sampling/sampling_logp_difference/max": 0.44519925117492676,
-      "sampling/sampling_logp_difference/mean": 0.09668958187103271,
+      "rewards/probe_shaping_dominance/mean": 0.96875,
+      "rewards/probe_shaping_dominance/std": 0.1767766922712326,
+      "rewards/probe_terminal_raw/mean": 0.03125,
+      "rewards/probe_terminal_raw/std": 0.1767766922712326,
+      "rewards/rollout_reward_func/mean": -0.9375,
+      "rewards/rollout_reward_func/std": 0.3535533845424652,
+      "sampling/importance_sampling_ratio/max": 1.9736851453781128,
+      "sampling/importance_sampling_ratio/mean": 0.8826993107795715,
+      "sampling/importance_sampling_ratio/min": 0.16903544962406158,
+      "sampling/sampling_logp_difference/max": 1.3518052101135254,
+      "sampling/sampling_logp_difference/mean": 0.09547412395477295,
       "step": 7,
-      "step_time": 11.251088656998036
+      "step_time": 10.242166022000674
     },
     {
-      "clip_ratio/high_max": 0.0,
-      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/high_max": 0.02083333395421505,
+      "clip_ratio/high_mean": 0.010416666977107525,
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.0,
-      "entropy": 1.006740152835846,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "entropy": 0.6963466256856918,
       "epoch": 8e-05,
-      "grad_norm": 3.747232403839007e-05,
-      "kl": 0.0050201611193188,
+      "grad_norm": 0.0007595560164190829,
+      "kl": 0.012291169143281877,
       "learning_rate": 1.6e-06,
       "loss": 0.0,
       "step": 8,
-      "step_time": 6.059573506999186
+      "step_time": 5.998965344997487
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -250,39 +250,39 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.90625,
-      "completions/mean_terminated_length": 2.90625,
-      "completions/min_length": 2.0,
-      "completions/min_terminated_length": 2.0,
-      "entropy": 0.6934512332081795,
+      "completions/mean_length": 3.0,
+      "completions/mean_terminated_length": 3.0,
+      "completions/min_length": 3.0,
+      "completions/min_terminated_length": 3.0,
+      "entropy": 0.6401678249239922,
       "epoch": 9e-05,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 3.45512708008755e-05,
-      "kl": 0.005398767949372996,
+      "grad_norm": 2.222989860456437e-05,
+      "kl": 0.0027450696452433476,
       "learning_rate": 1.8285714285714284e-06,
       "loss": 0.0,
-      "num_tokens": 226316.0,
-      "reward": 2.356250047683716,
-      "reward_std": 0.9791166186332703,
+      "num_tokens": 238781.0,
+      "reward": 1.9500000476837158,
+      "reward_std": 0.0,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.90625,
-      "rewards/probe_completion_length/std": 0.2961445748806,
+      "rewards/probe_completion_length/mean": 2.0,
+      "rewards/probe_completion_length/std": 0.0,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.75,
-      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
-      "rewards/probe_terminal_raw/mean": 0.25,
-      "rewards/probe_terminal_raw/std": 0.4399413466453552,
-      "rewards/rollout_reward_func/mean": -0.5,
-      "rewards/rollout_reward_func/std": 0.8798826932907104,
-      "sampling/importance_sampling_ratio/max": 1.4297953844070435,
-      "sampling/importance_sampling_ratio/mean": 0.9057049751281738,
-      "sampling/importance_sampling_ratio/min": 0.47415363788604736,
-      "sampling/sampling_logp_difference/max": 0.6755725145339966,
-      "sampling/sampling_logp_difference/mean": 0.07361993193626404,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.2344462871551514,
+      "sampling/importance_sampling_ratio/mean": 0.8851175904273987,
+      "sampling/importance_sampling_ratio/min": 0.5345703959465027,
+      "sampling/sampling_logp_difference/max": 0.5086061358451843,
+      "sampling/sampling_logp_difference/mean": 0.06150691583752632,
       "step": 9,
-      "step_time": 9.991546349998316
+      "step_time": 10.399385558001086
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -290,14 +290,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.6940568909049034,
+      "entropy": 0.6489881873130798,
       "epoch": 0.0001,
-      "grad_norm": 2.127706648025196e-05,
-      "kl": 0.0029271601816276416,
+      "grad_norm": 5.47774470760487e-05,
+      "kl": 0.005339258659660118,
       "learning_rate": 2.057142857142857e-06,
       "loss": 0.0,
       "step": 10,
-      "step_time": 6.432610926998677
+      "step_time": 6.213306221998209
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -312,35 +312,35 @@
       "completions/mean_terminated_length": 2.78125,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.6507774442434311,
+      "entropy": 0.6336817592382431,
       "epoch": 0.00011,
       "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.0026587501633912325,
-      "kl": 0.0030539491679064668,
+      "grad_norm": 0.0054480223916471004,
+      "kl": 0.0008345261200020104,
       "learning_rate": 2.2857142857142856e-06,
       "loss": -0.0,
-      "num_tokens": 269819.0,
-      "reward": 2.356250047683716,
-      "reward_std": 1.160071611404419,
+      "num_tokens": 281810.0,
+      "reward": 1.9187500476837158,
+      "reward_std": 0.7822372913360596,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
       "rewards/probe_completion_length/mean": 1.78125,
       "rewards/probe_completion_length/std": 0.420013427734375,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.6875,
-      "rewards/probe_shaping_dominance/std": 0.4709290862083435,
-      "rewards/probe_terminal_raw/mean": 0.3125,
-      "rewards/probe_terminal_raw/std": 0.4709290862083435,
-      "rewards/rollout_reward_func/mean": -0.375,
-      "rewards/rollout_reward_func/std": 0.941858172416687,
-      "sampling/importance_sampling_ratio/max": 1.1657392978668213,
-      "sampling/importance_sampling_ratio/mean": 0.9005649089813232,
-      "sampling/importance_sampling_ratio/min": 0.7608628273010254,
-      "sampling/sampling_logp_difference/max": 0.29668641090393066,
-      "sampling/sampling_logp_difference/mean": 0.05268556624650955,
+      "rewards/probe_shaping_dominance/mean": 0.90625,
+      "rewards/probe_shaping_dominance/std": 0.2961445748806,
+      "rewards/probe_terminal_raw/mean": 0.09375,
+      "rewards/probe_terminal_raw/std": 0.2961445748806,
+      "rewards/rollout_reward_func/mean": -0.8125,
+      "rewards/rollout_reward_func/std": 0.5922891497612,
+      "sampling/importance_sampling_ratio/max": 1.6787108182907104,
+      "sampling/importance_sampling_ratio/mean": 0.8936471939086914,
+      "sampling/importance_sampling_ratio/min": 0.5765784978866577,
+      "sampling/sampling_logp_difference/max": 0.5861668586730957,
+      "sampling/sampling_logp_difference/mean": 0.06395299732685089,
       "step": 11,
-      "step_time": 10.194988582000406
+      "step_time": 10.046140308000759
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -348,14 +348,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.6534814164042473,
+      "entropy": 0.6370653584599495,
       "epoch": 0.00012,
-      "grad_norm": 0.0026218171697109938,
-      "kl": 0.0028509022927494243,
+      "grad_norm": 0.005797144025564194,
+      "kl": 0.0016152820803654322,
       "learning_rate": 2.5142857142857142e-06,
       "loss": -0.0,
       "step": 12,
-      "step_time": 5.727766669999255
+      "step_time": 5.868836872001339
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -366,39 +366,39 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.59375,
-      "completions/mean_terminated_length": 2.59375,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.9608507007360458,
+      "entropy": 0.6840376108884811,
       "epoch": 0.00013,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 6.196741742314771e-05,
-      "kl": 0.006144683692127728,
+      "grad_norm": 2.0243625840521418e-05,
+      "kl": 0.0036721347541970317,
       "learning_rate": 2.742857142857143e-06,
       "loss": 0.0,
-      "num_tokens": 315178.0,
-      "reward": 2.043750047683716,
-      "reward_std": 1.201058030128479,
+      "num_tokens": 329969.0,
+      "reward": 1.7312500476837158,
+      "reward_std": 0.420013427734375,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.59375,
-      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_completion_length/mean": 1.78125,
+      "rewards/probe_completion_length/std": 0.420013427734375,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.75,
-      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
-      "rewards/probe_terminal_raw/mean": 0.25,
-      "rewards/probe_terminal_raw/std": 0.4399413466453552,
-      "rewards/rollout_reward_func/mean": -0.5,
-      "rewards/rollout_reward_func/std": 0.8798826932907104,
-      "sampling/importance_sampling_ratio/max": 1.196762204170227,
-      "sampling/importance_sampling_ratio/mean": 0.86090487241745,
-      "sampling/importance_sampling_ratio/min": 0.4875930845737457,
-      "sampling/sampling_logp_difference/max": 0.40385901927948,
-      "sampling/sampling_logp_difference/mean": 0.09211455285549164,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.2268524169921875,
+      "sampling/importance_sampling_ratio/mean": 0.9003363847732544,
+      "sampling/importance_sampling_ratio/min": 0.5451527833938599,
+      "sampling/sampling_logp_difference/max": 0.5136758089065552,
+      "sampling/sampling_logp_difference/mean": 0.06247905641794205,
       "step": 13,
-      "step_time": 10.002985979000186
+      "step_time": 10.74419506999584
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -406,72 +406,72 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.9843811690807343,
+      "entropy": 0.6921080499887466,
       "epoch": 0.00014,
-      "grad_norm": 9.995475556934252e-05,
-      "kl": 0.0112989693225245,
+      "grad_norm": 6.911743548698723e-05,
+      "kl": 0.007093713042195304,
       "learning_rate": 2.9714285714285716e-06,
       "loss": 0.0,
       "step": 14,
-      "step_time": 6.480481466999663
+      "step_time": 6.484958697003094
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.02083333395421505,
+      "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.02083333395421505,
+      "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.375,
-      "completions/mean_terminated_length": 2.375,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.8333694264292717,
+      "entropy": 0.7193514481186867,
       "epoch": 0.00015,
-      "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.009402290917932987,
-      "kl": 0.003834348764939932,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 1.1802055269072298e-05,
+      "kl": 0.001129702115576947,
       "learning_rate": 3.2e-06,
       "loss": 0.0,
-      "num_tokens": 360159.0,
-      "reward": 1.3875000476837158,
-      "reward_std": 0.6690146923065186,
+      "num_tokens": 378690.0,
+      "reward": 1.7312500476837158,
+      "reward_std": 0.420013427734375,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.375,
-      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_completion_length/mean": 1.78125,
+      "rewards/probe_completion_length/std": 0.420013427734375,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.96875,
-      "rewards/probe_shaping_dominance/std": 0.1767766922712326,
-      "rewards/probe_terminal_raw/mean": 0.03125,
-      "rewards/probe_terminal_raw/std": 0.1767766922712326,
-      "rewards/rollout_reward_func/mean": -0.9375,
-      "rewards/rollout_reward_func/std": 0.3535533845424652,
-      "sampling/importance_sampling_ratio/max": 2.0241074562072754,
-      "sampling/importance_sampling_ratio/mean": 0.9765468835830688,
-      "sampling/importance_sampling_ratio/min": 0.40592315793037415,
-      "sampling/sampling_logp_difference/max": 0.8091045618057251,
-      "sampling/sampling_logp_difference/mean": 0.09295991063117981,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.6583240032196045,
+      "sampling/importance_sampling_ratio/mean": 0.9662915468215942,
+      "sampling/importance_sampling_ratio/min": 0.763291597366333,
+      "sampling/sampling_logp_difference/max": 0.6004593372344971,
+      "sampling/sampling_logp_difference/mean": 0.06859122216701508,
       "step": 15,
-      "step_time": 10.550640546000068
+      "step_time": 10.628934043001209
     },
     {
-      "clip_ratio/high_max": 0.02083333395421505,
-      "clip_ratio/high_mean": 0.010416666977107525,
-      "clip_ratio/low_mean": 0.012500000186264515,
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.02291666716337204,
-      "entropy": 0.8196478858590126,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7239196375012398,
       "epoch": 0.00016,
-      "grad_norm": 0.001830586465075612,
-      "kl": 0.008265510034107137,
+      "grad_norm": 1.3546254194807261e-05,
+      "kl": 0.001586953027981508,
       "learning_rate": 3.428571428571428e-06,
       "loss": 0.0,
       "step": 16,
-      "step_time": 5.979319029999715
+      "step_time": 5.975834984001267
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -480,41 +480,41 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 8.0,
-      "completions/max_terminated_length": 8.0,
-      "completions/mean_length": 2.875,
-      "completions/mean_terminated_length": 2.875,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.46875,
+      "completions/mean_terminated_length": 2.46875,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.7344488576054573,
+      "entropy": 0.9238218814134598,
       "epoch": 0.00017,
-      "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.004971999675035477,
-      "kl": 0.00169508293402032,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 3.5807555832434446e-05,
+      "kl": 0.004726857401692541,
       "learning_rate": 3.657142857142857e-06,
       "loss": 0.0,
-      "num_tokens": 402339.0,
-      "reward": 2.2624998092651367,
-      "reward_std": 1.59510338306427,
+      "num_tokens": 426711.0,
+      "reward": 1.4187500476837158,
+      "reward_std": 0.507007360458374,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.9375,
-      "rewards/probe_completion_length/std": 1.3663583993911743,
+      "rewards/probe_completion_length/mean": 1.46875,
+      "rewards/probe_completion_length/std": 0.507007360458374,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.8125,
-      "rewards/probe_shaping_dominance/std": 0.3965577781200409,
-      "rewards/probe_terminal_raw/mean": 0.1875,
-      "rewards/probe_terminal_raw/std": 0.3965577781200409,
-      "rewards/rollout_reward_func/mean": -0.625,
-      "rewards/rollout_reward_func/std": 0.7931155562400818,
-      "sampling/importance_sampling_ratio/max": 1.4009569883346558,
-      "sampling/importance_sampling_ratio/mean": 0.8987216353416443,
-      "sampling/importance_sampling_ratio/min": 0.4638045132160187,
-      "sampling/sampling_logp_difference/max": 0.7497167587280273,
-      "sampling/sampling_logp_difference/mean": 0.08023187518119812,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.2269541025161743,
+      "sampling/importance_sampling_ratio/mean": 0.8787311315536499,
+      "sampling/importance_sampling_ratio/min": 0.5107928514480591,
+      "sampling/sampling_logp_difference/max": 0.5426488518714905,
+      "sampling/sampling_logp_difference/mean": 0.08957700431346893,
       "step": 17,
-      "step_time": 10.03037292999852
+      "step_time": 10.636119067996333
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -522,14 +522,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.7358419224619865,
+      "entropy": 0.9311410710215569,
       "epoch": 0.00018,
-      "grad_norm": 0.005262928549200296,
-      "kl": 0.0035523743881640257,
+      "grad_norm": 4.843952410737984e-05,
+      "kl": 0.0039593383044120856,
       "learning_rate": 3.885714285714286e-06,
       "loss": 0.0,
       "step": 18,
-      "step_time": 6.99632274600026
+      "step_time": 6.45704991400271
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -540,39 +540,39 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.75,
-      "completions/mean_terminated_length": 2.75,
+      "completions/mean_length": 2.40625,
+      "completions/mean_terminated_length": 2.40625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.6966657862067223,
+      "entropy": 0.9205264896154404,
       "epoch": 0.00019,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 0.0001658419641898945,
-      "kl": 0.020315399326136685,
+      "grad_norm": 2.6991318009095266e-05,
+      "kl": 0.005098032517707907,
       "learning_rate": 4.114285714285714e-06,
       "loss": 0.0,
-      "num_tokens": 452253.0,
-      "reward": 1.7000000476837158,
-      "reward_std": 0.4399413466453552,
+      "num_tokens": 476260.0,
+      "reward": 1.3875000476837158,
+      "reward_std": 0.504016101360321,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.75,
-      "rewards/probe_completion_length/std": 0.4399413466453552,
-      "rewards/probe_invalid_count/mean": 0.0,
-      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.40625,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
       "rewards/probe_shaping_dominance/mean": 1.0,
       "rewards/probe_shaping_dominance/std": 0.0,
       "rewards/probe_terminal_raw/mean": 0.0,
       "rewards/probe_terminal_raw/std": 0.0,
       "rewards/rollout_reward_func/mean": -1.0,
       "rewards/rollout_reward_func/std": 0.0,
-      "sampling/importance_sampling_ratio/max": 1.3721214532852173,
-      "sampling/importance_sampling_ratio/mean": 0.857242226600647,
-      "sampling/importance_sampling_ratio/min": 0.32355910539627075,
-      "sampling/sampling_logp_difference/max": 1.2377636432647705,
-      "sampling/sampling_logp_difference/mean": 0.09668726474046707,
+      "sampling/importance_sampling_ratio/max": 1.1448806524276733,
+      "sampling/importance_sampling_ratio/mean": 0.873583197593689,
+      "sampling/importance_sampling_ratio/min": 0.5131767988204956,
+      "sampling/sampling_logp_difference/max": 0.6054301261901855,
+      "sampling/sampling_logp_difference/mean": 0.08943168818950653,
       "step": 19,
-      "step_time": 10.249966756000504
+      "step_time": 10.410437912003545
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -580,14 +580,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.7012157067656517,
+      "entropy": 0.9188544452190399,
       "epoch": 0.0002,
-      "grad_norm": 2.9285800337675028e-05,
-      "kl": 0.003729396459675627,
+      "grad_norm": 2.9249089493532665e-05,
+      "kl": 0.00348761334316805,
       "learning_rate": 4.342857142857142e-06,
       "loss": 0.0,
       "step": 20,
-      "step_time": 6.159007310000561
+      "step_time": 6.143518736002079
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -596,56 +596,56 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 8.0,
-      "completions/max_terminated_length": 8.0,
-      "completions/mean_length": 2.875,
-      "completions/mean_terminated_length": 2.875,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.84375,
+      "completions/mean_terminated_length": 2.84375,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.8007383048534393,
+      "entropy": 0.6722570210695267,
       "epoch": 0.00021,
-      "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.007845139130949974,
-      "kl": 0.00453781802980302,
+      "frac_reward_zero_std": 0.5,
+      "grad_norm": 0.01012302003800869,
+      "kl": 0.009901609308144543,
       "learning_rate": 4.571428571428571e-06,
-      "loss": 0.0,
-      "num_tokens": 498950.0,
-      "reward": 2.1687498092651367,
-      "reward_std": 1.4308664798736572,
+      "loss": -0.0,
+      "num_tokens": 519927.0,
+      "reward": 2.231250047683716,
+      "reward_std": 0.9913944005966187,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.90625,
-      "rewards/probe_completion_length/std": 1.201058030128479,
+      "rewards/probe_completion_length/mean": 1.84375,
+      "rewards/probe_completion_length/std": 0.3689020276069641,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.84375,
-      "rewards/probe_shaping_dominance/std": 0.3689020276069641,
-      "rewards/probe_terminal_raw/mean": 0.15625,
-      "rewards/probe_terminal_raw/std": 0.3689020276069641,
-      "rewards/rollout_reward_func/mean": -0.6875,
-      "rewards/rollout_reward_func/std": 0.7378040552139282,
-      "sampling/importance_sampling_ratio/max": 1.7632538080215454,
-      "sampling/importance_sampling_ratio/mean": 0.9100350141525269,
-      "sampling/importance_sampling_ratio/min": 0.24061019718647003,
-      "sampling/sampling_logp_difference/max": 1.3275805711746216,
-      "sampling/sampling_logp_difference/mean": 0.08635358512401581,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.2477000951766968,
+      "sampling/importance_sampling_ratio/mean": 0.9216253161430359,
+      "sampling/importance_sampling_ratio/min": 0.6353219151496887,
+      "sampling/sampling_logp_difference/max": 0.462948203086853,
+      "sampling/sampling_logp_difference/mean": 0.0612877756357193,
       "step": 21,
-      "step_time": 9.962191802000234
+      "step_time": 10.465017245000126
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_mean": 0.012500000186264515,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.010416666977107525,
-      "entropy": 0.8003072440624237,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.6736566424369812,
       "epoch": 0.00022,
-      "grad_norm": 0.005664248485118151,
-      "kl": 0.008983593537777779,
+      "grad_norm": 0.007845910266041756,
+      "kl": 0.01028646872509853,
       "learning_rate": 4.8e-06,
-      "loss": 0.0,
+      "loss": -0.0,
       "step": 22,
-      "step_time": 6.554617122002128
+      "step_time": 6.385594672996376
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -656,54 +656,54 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 3.0,
-      "completions/mean_terminated_length": 3.0,
-      "completions/min_length": 3.0,
-      "completions/min_terminated_length": 3.0,
-      "entropy": 0.6100832298398018,
+      "completions/mean_length": 2.6875,
+      "completions/mean_terminated_length": 2.6875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8278574831783772,
       "epoch": 0.00023,
-      "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.007147330790758133,
-      "kl": 0.011804845751612447,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 5.582312951446511e-05,
+      "kl": 0.005553377578507934,
       "learning_rate": 5.0285714285714285e-06,
-      "loss": -0.0,
-      "num_tokens": 546931.0,
-      "reward": 2.325000047683716,
-      "reward_std": 0.7931155562400818,
+      "loss": 0.0,
+      "num_tokens": 568175.0,
+      "reward": 1.6375000476837158,
+      "reward_std": 0.4709290862083435,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 2.0,
-      "rewards/probe_completion_length/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.6875,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.8125,
-      "rewards/probe_shaping_dominance/std": 0.3965577781200409,
-      "rewards/probe_terminal_raw/mean": 0.1875,
-      "rewards/probe_terminal_raw/std": 0.3965577781200409,
-      "rewards/rollout_reward_func/mean": -0.625,
-      "rewards/rollout_reward_func/std": 0.7931155562400818,
-      "sampling/importance_sampling_ratio/max": 1.5701907873153687,
-      "sampling/importance_sampling_ratio/mean": 0.9152559041976929,
-      "sampling/importance_sampling_ratio/min": 0.38235926628112793,
-      "sampling/sampling_logp_difference/max": 0.8201756477355957,
-      "sampling/sampling_logp_difference/mean": 0.06971260905265808,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.3277029991149902,
+      "sampling/importance_sampling_ratio/mean": 0.8968680500984192,
+      "sampling/importance_sampling_ratio/min": 0.6297081708908081,
+      "sampling/sampling_logp_difference/max": 0.6801514625549316,
+      "sampling/sampling_logp_difference/mean": 0.08738863468170166,
       "step": 23,
-      "step_time": 9.90423563099921
+      "step_time": 10.24965080399852
     },
     {
-      "clip_ratio/high_max": 0.02083333395421505,
-      "clip_ratio/high_mean": 0.010416666977107525,
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.010416666977107525,
-      "entropy": 0.6199251338839531,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8206844255328178,
       "epoch": 0.00024,
-      "grad_norm": 0.005858343094587326,
-      "kl": 0.016530382781638764,
+      "grad_norm": 3.3745935070328414e-05,
+      "kl": 0.004957256046054681,
       "learning_rate": 5.257142857142857e-06,
-      "loss": -0.0,
+      "loss": 0.0,
       "step": 24,
-      "step_time": 5.902607746000285
+      "step_time": 6.033321787002933
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -714,39 +714,39 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.84375,
-      "completions/mean_terminated_length": 2.84375,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.5921763628721237,
+      "entropy": 0.6828285604715347,
       "epoch": 0.00025,
-      "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.0020523809362202883,
-      "kl": 0.002616291645608726,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 3.153899888275191e-05,
+      "kl": 0.0025460530912368995,
       "learning_rate": 5.485714285714286e-06,
       "loss": 0.0,
-      "num_tokens": 592263.0,
-      "reward": 2.731250047683716,
-      "reward_std": 1.2110878229141235,
+      "num_tokens": 616476.0,
+      "reward": 1.7312500476837158,
+      "reward_std": 0.420013427734375,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.84375,
-      "rewards/probe_completion_length/std": 0.3689020276069641,
+      "rewards/probe_completion_length/mean": 1.78125,
+      "rewards/probe_completion_length/std": 0.420013427734375,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.53125,
-      "rewards/probe_shaping_dominance/std": 0.507007360458374,
-      "rewards/probe_terminal_raw/mean": 0.46875,
-      "rewards/probe_terminal_raw/std": 0.507007360458374,
-      "rewards/rollout_reward_func/mean": -0.0625,
-      "rewards/rollout_reward_func/std": 1.014014720916748,
-      "sampling/importance_sampling_ratio/max": 1.470768690109253,
-      "sampling/importance_sampling_ratio/mean": 0.9247093200683594,
-      "sampling/importance_sampling_ratio/min": 0.46212413907051086,
-      "sampling/sampling_logp_difference/max": 0.4592486619949341,
-      "sampling/sampling_logp_difference/mean": 0.06738743931055069,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.0664699077606201,
+      "sampling/importance_sampling_ratio/mean": 0.8564985394477844,
+      "sampling/importance_sampling_ratio/min": 0.5391273498535156,
+      "sampling/sampling_logp_difference/max": 0.40851259231567383,
+      "sampling/sampling_logp_difference/mean": 0.06989122927188873,
       "step": 25,
-      "step_time": 10.45240628099964
+      "step_time": 10.793845786996826
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -754,14 +754,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.5830045342445374,
+      "entropy": 0.683635301887989,
       "epoch": 0.00026,
-      "grad_norm": 0.0027421230915933847,
-      "kl": 0.010621169898968219,
+      "grad_norm": 0.0001063640447682701,
+      "kl": 0.012343363827312714,
       "learning_rate": 5.7142857142857145e-06,
       "loss": 0.0,
       "step": 26,
-      "step_time": 5.986171178999939
+      "step_time": 6.026963442000124
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -772,54 +772,54 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 3.0,
-      "completions/mean_terminated_length": 3.0,
-      "completions/min_length": 3.0,
-      "completions/min_terminated_length": 3.0,
-      "entropy": 0.5842774957418442,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9264916479587555,
       "epoch": 0.00027,
-      "frac_reward_zero_std": 0.5,
-      "grad_norm": 0.00840095616877079,
-      "kl": 0.05193354899529368,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 8.221685857279226e-05,
+      "kl": 0.01160298454351505,
       "learning_rate": 5.942857142857143e-06,
       "loss": 0.0,
-      "num_tokens": 636215.0,
-      "reward": 2.575000047683716,
-      "reward_std": 0.941858172416687,
+      "num_tokens": 663041.0,
+      "reward": 1.5750000476837158,
+      "reward_std": 0.49186936020851135,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 2.0,
-      "rewards/probe_completion_length/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.6875,
-      "rewards/probe_shaping_dominance/std": 0.4709290862083435,
-      "rewards/probe_terminal_raw/mean": 0.3125,
-      "rewards/probe_terminal_raw/std": 0.4709290862083435,
-      "rewards/rollout_reward_func/mean": -0.375,
-      "rewards/rollout_reward_func/std": 0.941858172416687,
-      "sampling/importance_sampling_ratio/max": 1.1518341302871704,
-      "sampling/importance_sampling_ratio/mean": 0.9229466915130615,
-      "sampling/importance_sampling_ratio/min": 0.740210235118866,
-      "sampling/sampling_logp_difference/max": 0.23350971937179565,
-      "sampling/sampling_logp_difference/mean": 0.04789143055677414,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.4550175666809082,
+      "sampling/importance_sampling_ratio/mean": 0.8919216394424438,
+      "sampling/importance_sampling_ratio/min": 0.2783103287220001,
+      "sampling/sampling_logp_difference/max": 0.7940692901611328,
+      "sampling/sampling_logp_difference/mean": 0.09577842056751251,
       "step": 27,
-      "step_time": 9.939240472000165
+      "step_time": 10.399097397999867
     },
     {
-      "clip_ratio/high_max": 0.0416666679084301,
-      "clip_ratio/high_mean": 0.02083333395421505,
-      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.031250000931322575,
-      "entropy": 0.5928279720246792,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9363134279847145,
       "epoch": 0.00028,
-      "grad_norm": 0.001339235925115645,
-      "kl": 0.07154650982192834,
+      "grad_norm": 0.00021746440324932337,
+      "kl": 0.038634598463204384,
       "learning_rate": 6.171428571428571e-06,
-      "loss": -0.0,
+      "loss": 0.0,
       "step": 28,
-      "step_time": 5.932902072002435
+      "step_time": 6.575508214997171
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -828,41 +828,41 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 4.0,
-      "completions/max_terminated_length": 4.0,
-      "completions/mean_length": 2.34375,
-      "completions/mean_terminated_length": 2.34375,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.3125,
+      "completions/mean_terminated_length": 2.3125,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.9780827686190605,
+      "entropy": 0.9583756104111671,
       "epoch": 0.00029,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 6.672603194601834e-05,
-      "kl": 0.013599209080439323,
+      "grad_norm": 0.00010810867388499901,
+      "kl": 0.024504184257239103,
       "learning_rate": 6.4e-06,
       "loss": 0.0,
-      "num_tokens": 688661.0,
-      "reward": 1.3562500476837158,
-      "reward_std": 0.6652370095252991,
+      "num_tokens": 714798.0,
+      "reward": 1.2625000476837158,
+      "reward_std": 0.4709290862083435,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.34375,
-      "rewards/probe_completion_length/std": 0.5453246831893921,
-      "rewards/probe_invalid_count/mean": 0.0625,
-      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_completion_length/mean": 1.3125,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
       "rewards/probe_shaping_dominance/mean": 1.0,
       "rewards/probe_shaping_dominance/std": 0.0,
       "rewards/probe_terminal_raw/mean": 0.0,
       "rewards/probe_terminal_raw/std": 0.0,
       "rewards/rollout_reward_func/mean": -1.0,
       "rewards/rollout_reward_func/std": 0.0,
-      "sampling/importance_sampling_ratio/max": 1.0643420219421387,
-      "sampling/importance_sampling_ratio/mean": 0.8627969026565552,
-      "sampling/importance_sampling_ratio/min": 0.5205104351043701,
-      "sampling/sampling_logp_difference/max": 0.4734225273132324,
-      "sampling/sampling_logp_difference/mean": 0.0874757170677185,
+      "sampling/importance_sampling_ratio/max": 1.5010457038879395,
+      "sampling/importance_sampling_ratio/mean": 0.9024466276168823,
+      "sampling/importance_sampling_ratio/min": 0.24003513157367706,
+      "sampling/sampling_logp_difference/max": 1.3649718761444092,
+      "sampling/sampling_logp_difference/mean": 0.11732985079288483,
       "step": 29,
-      "step_time": 10.943690365000293
+      "step_time": 10.90805224900032
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -870,14 +870,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.9700995087623596,
+      "entropy": 0.9661187306046486,
       "epoch": 0.0003,
-      "grad_norm": 8.524458826286718e-05,
-      "kl": 0.01646642265313858,
+      "grad_norm": 4.100173100596294e-05,
+      "kl": 0.005046289617894217,
       "learning_rate": 6.628571428571428e-06,
       "loss": 0.0,
       "step": 30,
-      "step_time": 6.289740224002344
+      "step_time": 6.139430252998864
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -886,56 +886,56 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 3.0,
-      "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.8125,
-      "completions/mean_terminated_length": 2.8125,
+      "completions/max_length": 8.0,
+      "completions/max_terminated_length": 8.0,
+      "completions/mean_length": 3.0,
+      "completions/mean_terminated_length": 3.0,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.5977483242750168,
+      "entropy": 0.6211910024285316,
       "epoch": 0.00031,
       "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.0054800924845039845,
-      "kl": 0.03657999965313863,
+      "grad_norm": 0.004734576214104891,
+      "kl": 0.021804270082157018,
       "learning_rate": 6.857142857142856e-06,
       "loss": 0.0,
-      "num_tokens": 734504.0,
-      "reward": 2.325000047683716,
-      "reward_std": 1.0998533964157104,
+      "num_tokens": 760640.0,
+      "reward": 2.0749998092651367,
+      "reward_std": 1.361924648284912,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.8125,
-      "rewards/probe_completion_length/std": 0.3965577781200409,
+      "rewards/probe_completion_length/mean": 2.0625,
+      "rewards/probe_completion_length/std": 1.3182954788208008,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.71875,
-      "rewards/probe_shaping_dominance/std": 0.45680341124534607,
-      "rewards/probe_terminal_raw/mean": 0.28125,
-      "rewards/probe_terminal_raw/std": 0.45680341124534607,
-      "rewards/rollout_reward_func/mean": -0.4375,
-      "rewards/rollout_reward_func/std": 0.9136068224906921,
-      "sampling/importance_sampling_ratio/max": 0.9752265214920044,
-      "sampling/importance_sampling_ratio/mean": 0.8748180866241455,
-      "sampling/importance_sampling_ratio/min": 0.5280144214630127,
-      "sampling/sampling_logp_difference/max": 0.5450412034988403,
-      "sampling/sampling_logp_difference/mean": 0.05546283721923828,
+      "rewards/probe_shaping_dominance/mean": 0.96875,
+      "rewards/probe_shaping_dominance/std": 0.1767766922712326,
+      "rewards/probe_terminal_raw/mean": 0.03125,
+      "rewards/probe_terminal_raw/std": 0.1767766922712326,
+      "rewards/rollout_reward_func/mean": -0.9375,
+      "rewards/rollout_reward_func/std": 0.3535533845424652,
+      "sampling/importance_sampling_ratio/max": 1.617272138595581,
+      "sampling/importance_sampling_ratio/mean": 0.938959002494812,
+      "sampling/importance_sampling_ratio/min": 0.7246662974357605,
+      "sampling/sampling_logp_difference/max": 0.5712299346923828,
+      "sampling/sampling_logp_difference/mean": 0.05821167677640915,
       "step": 31,
-      "step_time": 10.040044982999461
+      "step_time": 10.732313114996941
     },
     {
-      "clip_ratio/high_max": 0.02083333395421505,
-      "clip_ratio/high_mean": 0.010416666977107525,
-      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.012500000186264515,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.02083333395421505,
-      "entropy": 0.5946150161325932,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.6255892366170883,
       "epoch": 0.00032,
-      "grad_norm": 0.00047032409929670393,
-      "kl": 0.06455633977139996,
+      "grad_norm": 0.002519609173759818,
+      "kl": 0.03869383437631768,
       "learning_rate": 7.085714285714285e-06,
       "loss": 0.0,
       "step": 32,
-      "step_time": 6.533442915998421
+      "step_time": 6.090148466997562
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -946,24 +946,24 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.21875,
-      "completions/mean_terminated_length": 2.21875,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 1.0080943554639816,
+      "entropy": 0.8118936568498611,
       "epoch": 0.00033,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 3.053375257877633e-05,
-      "kl": 0.010350218828534707,
+      "grad_norm": 8.35604005260393e-05,
+      "kl": 0.018937013923277846,
       "learning_rate": 7.314285714285714e-06,
       "loss": 0.0,
-      "num_tokens": 784204.0,
-      "reward": 1.1687500476837158,
-      "reward_std": 0.4200134575366974,
+      "num_tokens": 809340.0,
+      "reward": 1.5750000476837158,
+      "reward_std": 0.49186936020851135,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.21875,
-      "rewards/probe_completion_length/std": 0.420013427734375,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
       "rewards/probe_shaping_dominance/mean": 1.0,
@@ -972,13 +972,13 @@
       "rewards/probe_terminal_raw/std": 0.0,
       "rewards/rollout_reward_func/mean": -1.0,
       "rewards/rollout_reward_func/std": 0.0,
-      "sampling/importance_sampling_ratio/max": 1.6532065868377686,
-      "sampling/importance_sampling_ratio/mean": 0.8955996036529541,
-      "sampling/importance_sampling_ratio/min": 0.555946409702301,
-      "sampling/sampling_logp_difference/max": 0.6041266918182373,
-      "sampling/sampling_logp_difference/mean": 0.10231932997703552,
+      "sampling/importance_sampling_ratio/max": 1.808742642402649,
+      "sampling/importance_sampling_ratio/mean": 0.8965403437614441,
+      "sampling/importance_sampling_ratio/min": 0.4695906937122345,
+      "sampling/sampling_logp_difference/max": 0.6713973879814148,
+      "sampling/sampling_logp_difference/mean": 0.08081598579883575,
       "step": 33,
-      "step_time": 10.570013555000514
+      "step_time": 11.075895337000475
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -986,188 +986,188 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 1.0019470378756523,
+      "entropy": 0.8101335242390633,
       "epoch": 0.00034,
-      "grad_norm": 2.8502141503849998e-05,
-      "kl": 0.007157980311603751,
+      "grad_norm": 5.702187627321109e-05,
+      "kl": 0.019410209737543482,
       "learning_rate": 7.542857142857142e-06,
       "loss": 0.0,
       "step": 34,
-      "step_time": 6.132537012999819
+      "step_time": 6.524634418001369
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.02291666716337204,
+      "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.02291666716337204,
+      "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 3.0,
-      "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.71875,
-      "completions/mean_terminated_length": 2.71875,
+      "completions/max_length": 8.0,
+      "completions/max_terminated_length": 8.0,
+      "completions/mean_length": 3.03125,
+      "completions/mean_terminated_length": 3.03125,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.7035567462444305,
+      "entropy": 0.6599159687757492,
       "epoch": 0.00035,
       "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.0019313328666612506,
-      "kl": 0.12925253533012437,
+      "grad_norm": 0.0038426010869443417,
+      "kl": 0.0753691944591992,
       "learning_rate": 7.771428571428572e-06,
-      "loss": -0.0,
-      "num_tokens": 830986.0,
-      "reward": 2.262500047683716,
-      "reward_std": 1.1482806205749512,
+      "loss": 0.0,
+      "num_tokens": 857176.0,
+      "reward": 2.1687498092651367,
+      "reward_std": 1.385043978691101,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.71875,
-      "rewards/probe_completion_length/std": 0.45680341124534607,
-      "rewards/probe_invalid_count/mean": 0.03125,
-      "rewards/probe_invalid_count/std": 0.1767766922712326,
-      "rewards/probe_shaping_dominance/mean": 0.71875,
-      "rewards/probe_shaping_dominance/std": 0.45680341124534607,
-      "rewards/probe_terminal_raw/mean": 0.28125,
-      "rewards/probe_terminal_raw/std": 0.45680341124534607,
-      "rewards/rollout_reward_func/mean": -0.4375,
-      "rewards/rollout_reward_func/std": 0.9136068224906921,
-      "sampling/importance_sampling_ratio/max": 1.0661519765853882,
-      "sampling/importance_sampling_ratio/mean": 0.8217430710792542,
-      "sampling/importance_sampling_ratio/min": 0.204800084233284,
-      "sampling/sampling_logp_difference/max": 1.4816503524780273,
-      "sampling/sampling_logp_difference/mean": 0.09921938180923462,
+      "rewards/probe_completion_length/mean": 2.09375,
+      "rewards/probe_completion_length/std": 1.304072380065918,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.9375,
+      "rewards/probe_shaping_dominance/std": 0.24593468010425568,
+      "rewards/probe_terminal_raw/mean": 0.0625,
+      "rewards/probe_terminal_raw/std": 0.24593468010425568,
+      "rewards/rollout_reward_func/mean": -0.875,
+      "rewards/rollout_reward_func/std": 0.49186936020851135,
+      "sampling/importance_sampling_ratio/max": 1.3117048740386963,
+      "sampling/importance_sampling_ratio/mean": 0.8893284201622009,
+      "sampling/importance_sampling_ratio/min": 0.4679924249649048,
+      "sampling/sampling_logp_difference/max": 0.5346801280975342,
+      "sampling/sampling_logp_difference/mean": 0.06689614057540894,
       "step": 35,
-      "step_time": 10.03484531599861
+      "step_time": 10.305036647003362
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.02291666716337204,
+      "clip_ratio/low_mean": 0.04375000111758709,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.02291666716337204,
-      "entropy": 0.7101162374019623,
+      "clip_ratio/region_mean": 0.04375000111758709,
+      "entropy": 0.6635555773973465,
       "epoch": 0.00036,
-      "grad_norm": 0.0020328587852418423,
-      "kl": 0.10064423059520777,
+      "grad_norm": 0.0011747012613341212,
+      "kl": 0.13147676514927298,
       "learning_rate": 8e-06,
       "loss": -0.0,
       "step": 36,
-      "step_time": 6.561449165999875
+      "step_time": 6.042067124002642
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.058333334513008595,
+      "clip_ratio/low_mean": 0.010416666977107525,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.058333334513008595,
+      "clip_ratio/region_mean": 0.010416666977107525,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 3.0,
-      "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.59375,
-      "completions/mean_terminated_length": 2.59375,
+      "completions/max_length": 6.0,
+      "completions/max_terminated_length": 6.0,
+      "completions/mean_length": 2.96875,
+      "completions/mean_terminated_length": 2.96875,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.8379025869071484,
+      "entropy": 0.6612287536263466,
       "epoch": 0.00037,
-      "frac_reward_zero_std": 0.5,
-      "grad_norm": 0.002047101268544793,
-      "kl": 0.4796946058049798,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0005710122059099376,
+      "kl": 0.27994170751958336,
       "learning_rate": 7.99999999962976e-06,
-      "loss": -0.0001,
-      "num_tokens": 878912.0,
-      "reward": 1.6687500476837158,
-      "reward_std": 0.7718588709831238,
+      "loss": 0.0,
+      "num_tokens": 900229.0,
+      "reward": 2.356250047683716,
+      "reward_std": 1.0734140872955322,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.59375,
-      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_completion_length/mean": 1.96875,
+      "rewards/probe_completion_length/std": 0.6467972993850708,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.9375,
-      "rewards/probe_shaping_dominance/std": 0.24593468010425568,
-      "rewards/probe_terminal_raw/mean": 0.0625,
-      "rewards/probe_terminal_raw/std": 0.24593468010425568,
-      "rewards/rollout_reward_func/mean": -0.875,
-      "rewards/rollout_reward_func/std": 0.49186936020851135,
-      "sampling/importance_sampling_ratio/max": 1.027855396270752,
-      "sampling/importance_sampling_ratio/mean": 0.7403821349143982,
-      "sampling/importance_sampling_ratio/min": 0.20094381272792816,
-      "sampling/sampling_logp_difference/max": 1.4957327842712402,
-      "sampling/sampling_logp_difference/mean": 0.15929864346981049,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.33085298538208,
+      "sampling/importance_sampling_ratio/mean": 0.9053265452384949,
+      "sampling/importance_sampling_ratio/min": 0.3018966317176819,
+      "sampling/sampling_logp_difference/max": 0.954264760017395,
+      "sampling/sampling_logp_difference/mean": 0.07190588116645813,
       "step": 37,
-      "step_time": 10.421861329000421
+      "step_time": 10.737445328000831
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.09166666865348816,
-      "clip_ratio/low_min": 0.02083333395421505,
-      "clip_ratio/region_mean": 0.09166666865348816,
-      "entropy": 0.8225769698619843,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6657588593661785,
       "epoch": 0.00038,
-      "grad_norm": 0.0023972741328179836,
-      "kl": 0.7458103867247701,
+      "grad_norm": 0.013408876024186611,
+      "kl": 0.4735182789929695,
       "learning_rate": 7.99999999851904e-06,
-      "loss": -0.0001,
+      "loss": 0.0001,
       "step": 38,
-      "step_time": 6.042584625999552
+      "step_time": 5.906282284999179
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.012500000186264515,
+      "clip_ratio/low_mean": 0.02083333395421505,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.012500000186264515,
+      "clip_ratio/region_mean": 0.02083333395421505,
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.6875,
-      "completions/mean_terminated_length": 2.6875,
+      "completions/mean_length": 2.90625,
+      "completions/mean_terminated_length": 2.90625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.7016490623354912,
+      "entropy": 0.6484945714473724,
       "epoch": 0.00039,
-      "frac_reward_zero_std": 0.5,
-      "grad_norm": 0.0051078652031719685,
-      "kl": 0.3725941587081252,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.004464410245418549,
+      "kl": 0.35673816324560903,
       "learning_rate": 7.999999996667841e-06,
-      "loss": -0.0002,
-      "num_tokens": 926402.0,
-      "reward": 1.8875000476837158,
-      "reward_std": 0.9136068224906921,
+      "loss": -0.0,
+      "num_tokens": 948183.0,
+      "reward": 2.075000047683716,
+      "reward_std": 0.6599119901657104,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.6875,
-      "rewards/probe_completion_length/std": 0.4709290862083435,
-      "rewards/probe_invalid_count/mean": 0.0,
-      "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.875,
-      "rewards/probe_shaping_dominance/std": 0.33601075410842896,
-      "rewards/probe_terminal_raw/mean": 0.125,
-      "rewards/probe_terminal_raw/std": 0.33601075410842896,
-      "rewards/rollout_reward_func/mean": -0.75,
-      "rewards/rollout_reward_func/std": 0.6720215082168579,
-      "sampling/importance_sampling_ratio/max": 2.5818870067596436,
-      "sampling/importance_sampling_ratio/mean": 0.9834462404251099,
-      "sampling/importance_sampling_ratio/min": 0.6739128828048706,
-      "sampling/sampling_logp_difference/max": 1.015700340270996,
-      "sampling/sampling_logp_difference/mean": 0.09071232378482819,
+      "rewards/probe_completion_length/mean": 1.90625,
+      "rewards/probe_completion_length/std": 0.2961445748806,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.90625,
+      "rewards/probe_shaping_dominance/std": 0.2961445748806,
+      "rewards/probe_terminal_raw/mean": 0.09375,
+      "rewards/probe_terminal_raw/std": 0.2961445748806,
+      "rewards/rollout_reward_func/mean": -0.8125,
+      "rewards/rollout_reward_func/std": 0.5922891497612,
+      "sampling/importance_sampling_ratio/max": 1.1827647686004639,
+      "sampling/importance_sampling_ratio/mean": 0.8997206687927246,
+      "sampling/importance_sampling_ratio/min": 0.4464786946773529,
+      "sampling/sampling_logp_difference/max": 0.7061750888824463,
+      "sampling/sampling_logp_difference/mean": 0.07010062038898468,
       "step": 39,
-      "step_time": 10.201964306000264
+      "step_time": 10.68057268800294
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.04375000111758709,
+      "clip_ratio/low_mean": 0.031250000931322575,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.04375000111758709,
-      "entropy": 0.7204578444361687,
+      "clip_ratio/region_mean": 0.031250000931322575,
+      "entropy": 0.6486207321286201,
       "epoch": 0.0004,
-      "grad_norm": 0.0035937828943133354,
-      "kl": 0.380961946604657,
+      "grad_norm": 0.0022209053859114647,
+      "kl": 0.46453470387496054,
       "learning_rate": 7.999999994076165e-06,
-      "loss": -0.0002,
+      "loss": -0.0,
       "step": 40,
-      "step_time": 6.1126418210005795
+      "step_time": 7.016006264999305
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1176,41 +1176,41 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 7.0,
-      "completions/max_terminated_length": 7.0,
-      "completions/mean_length": 2.59375,
-      "completions/mean_terminated_length": 2.59375,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.84375,
+      "completions/mean_terminated_length": 2.84375,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.7181754857301712,
+      "entropy": 0.5532779470086098,
       "epoch": 0.00041,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 0.0032158580143004656,
-      "kl": 0.5504793294239789,
+      "grad_norm": 4.3166332034161314e-05,
+      "kl": 0.08521343817847082,
       "learning_rate": 7.999999990744006e-06,
       "loss": 0.0,
-      "num_tokens": 976326.0,
-      "reward": 1.6062500476837158,
-      "reward_std": 1.260040283203125,
+      "num_tokens": 991447.0,
+      "reward": 3.293750047683716,
+      "reward_std": 1.180742621421814,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.65625,
-      "rewards/probe_completion_length/std": 1.260040283203125,
+      "rewards/probe_completion_length/mean": 1.84375,
+      "rewards/probe_completion_length/std": 0.3689020276069641,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 1.0,
-      "rewards/probe_shaping_dominance/std": 0.0,
-      "rewards/probe_terminal_raw/mean": 0.0,
-      "rewards/probe_terminal_raw/std": 0.0,
-      "rewards/rollout_reward_func/mean": -1.0,
-      "rewards/rollout_reward_func/std": 0.0,
-      "sampling/importance_sampling_ratio/max": 1.2834391593933105,
-      "sampling/importance_sampling_ratio/mean": 0.8834752440452576,
-      "sampling/importance_sampling_ratio/min": 0.009724746458232403,
-      "sampling/sampling_logp_difference/max": 4.535632610321045,
-      "sampling/sampling_logp_difference/mean": 0.11896064877510071,
+      "rewards/probe_shaping_dominance/mean": 0.25,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.75,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": 0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.0786104202270508,
+      "sampling/importance_sampling_ratio/mean": 0.9027760624885559,
+      "sampling/importance_sampling_ratio/min": 0.4814170300960541,
+      "sampling/sampling_logp_difference/max": 0.5579321980476379,
+      "sampling/sampling_logp_difference/mean": 0.04839656874537468,
       "step": 41,
-      "step_time": 10.60360760599906
+      "step_time": 10.06029906799813
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1218,79 +1218,79 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.7254802733659744,
+      "entropy": 0.5597705021500587,
       "epoch": 0.00042,
-      "grad_norm": 0.005343083292245865,
-      "kl": 0.9199199170143402,
+      "grad_norm": 7.465948147000745e-05,
+      "kl": 0.0919022122034221,
       "learning_rate": 7.999999986671369e-06,
       "loss": 0.0,
       "step": 42,
-      "step_time": 6.086681632998989
+      "step_time": 5.93805161900309
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.010416666977107525,
+      "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.6875,
-      "completions/mean_terminated_length": 2.6875,
+      "completions/mean_length": 2.5,
+      "completions/mean_terminated_length": 2.5,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.7109739929437637,
+      "entropy": 0.8421927690505981,
       "epoch": 0.00043,
-      "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.0019174222834408283,
-      "kl": 0.9979303739964962,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0004577457730192691,
+      "kl": 0.09166453240914052,
       "learning_rate": 7.999999981858253e-06,
-      "loss": -0.0001,
-      "num_tokens": 1024501.0,
-      "reward": 2.450000047683716,
-      "reward_std": 1.319823980331421,
+      "loss": 0.0,
+      "num_tokens": 1039971.0,
+      "reward": 1.4500000476837158,
+      "reward_std": 0.5080004930496216,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.6875,
-      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_completion_length/mean": 1.5,
+      "rewards/probe_completion_length/std": 0.5080004930496216,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.59375,
-      "rewards/probe_shaping_dominance/std": 0.49899089336395264,
-      "rewards/probe_terminal_raw/mean": 0.40625,
-      "rewards/probe_terminal_raw/std": 0.49899089336395264,
-      "rewards/rollout_reward_func/mean": -0.1875,
-      "rewards/rollout_reward_func/std": 0.9979817867279053,
-      "sampling/importance_sampling_ratio/max": 2.3034539222717285,
-      "sampling/importance_sampling_ratio/mean": 1.0141327381134033,
-      "sampling/importance_sampling_ratio/min": 0.10541930049657822,
-      "sampling/sampling_logp_difference/max": 2.156667470932007,
-      "sampling/sampling_logp_difference/mean": 0.1373995542526245,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.2432236671447754,
+      "sampling/importance_sampling_ratio/mean": 0.845447838306427,
+      "sampling/importance_sampling_ratio/min": 0.2577609121799469,
+      "sampling/sampling_logp_difference/max": 1.228003740310669,
+      "sampling/sampling_logp_difference/mean": 0.10384337604045868,
       "step": 43,
-      "step_time": 10.50327422799728
+      "step_time": 10.771515129998079
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.010416666977107525,
-      "entropy": 0.7114239484071732,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8766651302576065,
       "epoch": 0.00044,
-      "grad_norm": 0.0017295810393989086,
-      "kl": 1.1500265896320343,
+      "grad_norm": 0.0008605438051745296,
+      "kl": 0.1548362478843046,
       "learning_rate": 7.999999976304658e-06,
-      "loss": -0.0001,
+      "loss": 0.0,
       "step": 44,
-      "step_time": 6.055206838999766
+      "step_time": 6.052436936000959
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_mean": 0.012500000186264515,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.0,
+      "clip_ratio/region_mean": 0.012500000186264515,
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
@@ -1298,50 +1298,50 @@
       "completions/mean_terminated_length": 2.625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.6414694786071777,
+      "entropy": 0.651808388531208,
       "epoch": 0.00045,
       "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.0025684612337499857,
-      "kl": 0.405933300498873,
+      "grad_norm": 0.0012912065722048283,
+      "kl": 0.4913899615421542,
       "learning_rate": 7.999999970010581e-06,
       "loss": -0.0001,
-      "num_tokens": 1066496.0,
-      "reward": 2.512500047683716,
-      "reward_std": 1.4127871990203857,
+      "num_tokens": 1089788.0,
+      "reward": 2.012500047683716,
+      "reward_std": 1.1341474056243896,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
       "rewards/probe_completion_length/mean": 1.625,
       "rewards/probe_completion_length/std": 0.49186936020851135,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.53125,
-      "rewards/probe_shaping_dominance/std": 0.507007360458374,
-      "rewards/probe_terminal_raw/mean": 0.46875,
-      "rewards/probe_terminal_raw/std": 0.507007360458374,
-      "rewards/rollout_reward_func/mean": -0.0625,
-      "rewards/rollout_reward_func/std": 1.014014720916748,
-      "sampling/importance_sampling_ratio/max": 2.0344650745391846,
-      "sampling/importance_sampling_ratio/mean": 0.9185701608657837,
-      "sampling/importance_sampling_ratio/min": 0.3526281416416168,
-      "sampling/sampling_logp_difference/max": 0.9392699599266052,
-      "sampling/sampling_logp_difference/mean": 0.06703778356313705,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.964327096939087,
+      "sampling/importance_sampling_ratio/mean": 0.9429373741149902,
+      "sampling/importance_sampling_ratio/min": 0.45201072096824646,
+      "sampling/sampling_logp_difference/max": 0.7543458938598633,
+      "sampling/sampling_logp_difference/mean": 0.08115933835506439,
       "step": 45,
-      "step_time": 10.408880783999848
+      "step_time": 10.30191364899838
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_mean": 0.012500000186264515,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.010416666977107525,
-      "entropy": 0.6403752155601978,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.6510747969150543,
       "epoch": 0.00046,
-      "grad_norm": 0.0011193460086360574,
-      "kl": 0.418248422909528,
+      "grad_norm": 0.0015278910286724567,
+      "kl": 0.46892713010311127,
       "learning_rate": 7.999999962976027e-06,
       "loss": -0.0001,
       "step": 46,
-      "step_time": 5.940715425000235
+      "step_time": 6.073613009000837
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1352,39 +1352,39 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.75,
-      "completions/mean_terminated_length": 2.75,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.5939470268785954,
+      "entropy": 0.7168305143713951,
       "epoch": 0.00047,
-      "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.00029027138953097165,
-      "kl": 1.086835566908121,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 5.2330131438793615e-05,
+      "kl": 0.09791750832391699,
       "learning_rate": 7.999999955200991e-06,
-      "loss": -0.0,
-      "num_tokens": 1115763.0,
-      "reward": 2.137500047683716,
-      "reward_std": 1.0606601238250732,
+      "loss": 0.0,
+      "num_tokens": 1140406.0,
+      "reward": 1.5750000476837158,
+      "reward_std": 0.49186936020851135,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.75,
-      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.78125,
-      "rewards/probe_shaping_dominance/std": 0.420013427734375,
-      "rewards/probe_terminal_raw/mean": 0.21875,
-      "rewards/probe_terminal_raw/std": 0.420013427734375,
-      "rewards/rollout_reward_func/mean": -0.5625,
-      "rewards/rollout_reward_func/std": 0.84002685546875,
-      "sampling/importance_sampling_ratio/max": 1.3324123620986938,
-      "sampling/importance_sampling_ratio/mean": 0.921471357345581,
-      "sampling/importance_sampling_ratio/min": 0.3086128532886505,
-      "sampling/sampling_logp_difference/max": 1.1996221542358398,
-      "sampling/sampling_logp_difference/mean": 0.09517379105091095,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.1620327234268188,
+      "sampling/importance_sampling_ratio/mean": 0.8909596800804138,
+      "sampling/importance_sampling_ratio/min": 0.521927535533905,
+      "sampling/sampling_logp_difference/max": 0.5213262438774109,
+      "sampling/sampling_logp_difference/mean": 0.06901270151138306,
       "step": 47,
-      "step_time": 10.570694217000892
+      "step_time": 11.32904220099772
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1392,14 +1392,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.5917784459888935,
+      "entropy": 0.7062753215432167,
       "epoch": 0.00048,
-      "grad_norm": 0.00045079184928908944,
-      "kl": 1.1454527378082275,
+      "grad_norm": 5.481049811351113e-05,
+      "kl": 0.10177884640688717,
       "learning_rate": 7.999999946685478e-06,
-      "loss": -0.0,
+      "loss": 0.0,
       "step": 48,
-      "step_time": 6.084416185999544
+      "step_time": 6.08483760800118
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1408,41 +1408,41 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 3.0,
-      "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.6875,
-      "completions/mean_terminated_length": 2.6875,
+      "completions/max_length": 7.0,
+      "completions/max_terminated_length": 7.0,
+      "completions/mean_length": 2.75,
+      "completions/mean_terminated_length": 2.75,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.5497034415602684,
+      "entropy": 0.7744219973683357,
       "epoch": 0.00049,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 4.048119080835022e-05,
-      "kl": 1.0796112641692162,
+      "grad_norm": 0.00023119112302083522,
+      "kl": 0.05695398036914412,
       "learning_rate": 7.999999937429484e-06,
       "loss": 0.0,
-      "num_tokens": 1159887.0,
-      "reward": 2.637500047683716,
-      "reward_std": 1.3781123161315918,
+      "num_tokens": 1184107.0,
+      "reward": 2.262500047683716,
+      "reward_std": 1.574750304222107,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.6875,
-      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_completion_length/mean": 1.8125,
+      "rewards/probe_completion_length/std": 1.2296733856201172,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.5,
-      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
-      "rewards/probe_terminal_raw/mean": 0.5,
-      "rewards/probe_terminal_raw/std": 0.5080004930496216,
-      "rewards/rollout_reward_func/mean": 0.0,
-      "rewards/rollout_reward_func/std": 1.0160009860992432,
-      "sampling/importance_sampling_ratio/max": 1.194502592086792,
-      "sampling/importance_sampling_ratio/mean": 0.8948264122009277,
-      "sampling/importance_sampling_ratio/min": 0.0,
-      "sampling/sampling_logp_difference/max": 1.9638302326202393,
-      "sampling/sampling_logp_difference/mean": 0.07986211776733398,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.0071394443511963,
+      "sampling/importance_sampling_ratio/mean": 0.8390346169471741,
+      "sampling/importance_sampling_ratio/min": 0.4063165485858917,
+      "sampling/sampling_logp_difference/max": 0.7753506898880005,
+      "sampling/sampling_logp_difference/mean": 0.0786946639418602,
       "step": 49,
-      "step_time": 10.586019360999671
+      "step_time": 10.03044281899929
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1450,57 +1450,57 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.5474356599152088,
+      "entropy": 0.7632701024413109,
       "epoch": 0.0005,
-      "grad_norm": 4.84514057461638e-05,
-      "kl": 1.0822610259056091,
+      "grad_norm": 0.0003675088519230485,
+      "kl": 0.08831967870355584,
       "learning_rate": 7.999999927433012e-06,
       "loss": 0.0,
       "step": 50,
-      "step_time": 6.64471124299871
+      "step_time": 6.67380327900355
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.625,
-      "completions/mean_terminated_length": 2.625,
+      "completions/mean_length": 2.75,
+      "completions/mean_terminated_length": 2.75,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.6383445039391518,
+      "entropy": 0.883178424090147,
       "epoch": 0.00051,
-      "frac_reward_zero_std": 1.0,
-      "grad_norm": 0.00015050763613544405,
-      "kl": 0.6388462577888276,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.004802894778549671,
+      "kl": 1.0498504750430584,
       "learning_rate": 7.99999991669606e-06,
       "loss": 0.0,
-      "num_tokens": 1206661.0,
-      "reward": 2.075000047683716,
-      "reward_std": 1.1845782995224,
+      "num_tokens": 1231279.0,
+      "reward": 1.8250000476837158,
+      "reward_std": 0.7071067690849304,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.625,
-      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_completion_length/mean": 1.75,
+      "rewards/probe_completion_length/std": 0.4399413466453552,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.75,
-      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
-      "rewards/probe_terminal_raw/mean": 0.25,
-      "rewards/probe_terminal_raw/std": 0.4399413466453552,
-      "rewards/rollout_reward_func/mean": -0.5,
-      "rewards/rollout_reward_func/std": 0.8798826932907104,
-      "sampling/importance_sampling_ratio/max": 2.2058069705963135,
-      "sampling/importance_sampling_ratio/mean": 0.9211057424545288,
-      "sampling/importance_sampling_ratio/min": 0.5255441069602966,
-      "sampling/sampling_logp_difference/max": 0.8412450551986694,
-      "sampling/sampling_logp_difference/mean": 0.07937976717948914,
+      "rewards/probe_shaping_dominance/mean": 0.9375,
+      "rewards/probe_shaping_dominance/std": 0.24593468010425568,
+      "rewards/probe_terminal_raw/mean": 0.0625,
+      "rewards/probe_terminal_raw/std": 0.24593468010425568,
+      "rewards/rollout_reward_func/mean": -0.875,
+      "rewards/rollout_reward_func/std": 0.49186936020851135,
+      "sampling/importance_sampling_ratio/max": 1.3820379972457886,
+      "sampling/importance_sampling_ratio/mean": 0.6330314874649048,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 3.0561718940734863,
+      "sampling/sampling_logp_difference/mean": 0.28779780864715576,
       "step": 51,
-      "step_time": 10.090468297998086
+      "step_time": 10.563791698999921
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1508,14 +1508,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.65125298127532,
+      "entropy": 0.8849975541234016,
       "epoch": 0.00052,
-      "grad_norm": 0.00012597418390214443,
-      "kl": 0.6565363313129637,
+      "grad_norm": 0.0024871043860912323,
+      "kl": 0.7585216974839568,
       "learning_rate": 7.999999905218627e-06,
-      "loss": 0.0,
+      "loss": -0.0,
       "step": 52,
-      "step_time": 6.536847638998552
+      "step_time": 6.215853247000268
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1526,112 +1526,112 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.875,
-      "completions/mean_terminated_length": 2.875,
+      "completions/mean_length": 2.46875,
+      "completions/mean_terminated_length": 2.46875,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.5926927104592323,
+      "entropy": 0.8856398165225983,
       "epoch": 0.00053,
-      "frac_reward_zero_std": 1.0,
-      "grad_norm": 0.00019273272482678294,
-      "kl": 0.5490526768262498,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0032466028351336718,
+      "kl": 0.22195444721728563,
       "learning_rate": 7.999999893000716e-06,
-      "loss": 0.0,
-      "num_tokens": 1247965.0,
-      "reward": 2.825000047683716,
-      "reward_std": 1.1845782995224,
+      "loss": -0.0,
+      "num_tokens": 1285108.0,
+      "reward": 1.5437500476837158,
+      "reward_std": 0.797551691532135,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.875,
-      "rewards/probe_completion_length/std": 0.33601075410842896,
+      "rewards/probe_completion_length/mean": 1.46875,
+      "rewards/probe_completion_length/std": 0.507007360458374,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.5,
-      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
-      "rewards/probe_terminal_raw/mean": 0.5,
-      "rewards/probe_terminal_raw/std": 0.5080004930496216,
-      "rewards/rollout_reward_func/mean": 0.0,
-      "rewards/rollout_reward_func/std": 1.0160009860992432,
-      "sampling/importance_sampling_ratio/max": 1.0803444385528564,
-      "sampling/importance_sampling_ratio/mean": 0.8436357975006104,
-      "sampling/importance_sampling_ratio/min": 0.22206127643585205,
-      "sampling/sampling_logp_difference/max": 1.4294672012329102,
-      "sampling/sampling_logp_difference/mean": 0.08037947118282318,
+      "rewards/probe_shaping_dominance/mean": 0.9375,
+      "rewards/probe_shaping_dominance/std": 0.24593468010425568,
+      "rewards/probe_terminal_raw/mean": 0.0625,
+      "rewards/probe_terminal_raw/std": 0.24593468010425568,
+      "rewards/rollout_reward_func/mean": -0.875,
+      "rewards/rollout_reward_func/std": 0.49186936020851135,
+      "sampling/importance_sampling_ratio/max": 1.8507949113845825,
+      "sampling/importance_sampling_ratio/mean": 0.9760150909423828,
+      "sampling/importance_sampling_ratio/min": 0.4805029332637787,
+      "sampling/sampling_logp_difference/max": 0.6980669498443604,
+      "sampling/sampling_logp_difference/mean": 0.10727604478597641,
       "step": 53,
-      "step_time": 9.839941635998002
+      "step_time": 11.183175942000162
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.0,
-      "entropy": 0.571772389113903,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "entropy": 0.8866911977529526,
       "epoch": 0.00054,
-      "grad_norm": 0.0002839839144144207,
-      "kl": 0.5748571505537257,
+      "grad_norm": 0.0019594912882894278,
+      "kl": 0.21692284429445863,
       "learning_rate": 7.999999880042326e-06,
-      "loss": 0.0,
+      "loss": -0.0,
       "step": 54,
-      "step_time": 6.453985008000927
+      "step_time": 6.095816975001071
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.010416666977107525,
+      "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 3.0,
-      "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.71875,
-      "completions/mean_terminated_length": 2.71875,
+      "completions/max_length": 7.0,
+      "completions/max_terminated_length": 7.0,
+      "completions/mean_length": 2.8125,
+      "completions/mean_terminated_length": 2.8125,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.573297243565321,
+      "entropy": 0.8108381032943726,
       "epoch": 0.00055,
       "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.003776031779125333,
-      "kl": 1.215886102989316,
+      "grad_norm": 0.003048208076506853,
+      "kl": 0.43219664352363907,
       "learning_rate": 7.999999866343456e-06,
       "loss": -0.0,
-      "num_tokens": 1295187.0,
-      "reward": 1.7312500476837158,
-      "reward_std": 0.6082431077957153,
+      "num_tokens": 1334379.0,
+      "reward": 2.6374998092651367,
+      "reward_std": 1.63504958152771,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.71875,
-      "rewards/probe_completion_length/std": 0.45680341124534607,
+      "rewards/probe_completion_length/mean": 1.875,
+      "rewards/probe_completion_length/std": 1.2115039825439453,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.96875,
-      "rewards/probe_shaping_dominance/std": 0.1767766922712326,
-      "rewards/probe_terminal_raw/mean": 0.03125,
-      "rewards/probe_terminal_raw/std": 0.1767766922712326,
-      "rewards/rollout_reward_func/mean": -0.9375,
-      "rewards/rollout_reward_func/std": 0.3535533845424652,
-      "sampling/importance_sampling_ratio/max": 2.5861928462982178,
-      "sampling/importance_sampling_ratio/mean": 0.8224544525146484,
-      "sampling/importance_sampling_ratio/min": 0.0271281935274601,
-      "sampling/sampling_logp_difference/max": 3.4792683124542236,
-      "sampling/sampling_logp_difference/mean": 0.2531231939792633,
+      "rewards/probe_shaping_dominance/mean": 0.59375,
+      "rewards/probe_shaping_dominance/std": 0.49899089336395264,
+      "rewards/probe_terminal_raw/mean": 0.40625,
+      "rewards/probe_terminal_raw/std": 0.49899089336395264,
+      "rewards/rollout_reward_func/mean": -0.1875,
+      "rewards/rollout_reward_func/std": 0.9979817867279053,
+      "sampling/importance_sampling_ratio/max": 1.2705219984054565,
+      "sampling/importance_sampling_ratio/mean": 0.7683707475662231,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 1.6249732971191406,
+      "sampling/sampling_logp_difference/mean": 0.12966416776180267,
       "step": 55,
-      "step_time": 10.005798918000437
+      "step_time": 10.185927998001716
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.0,
-      "entropy": 0.579718828201294,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "entropy": 0.7804929316043854,
       "epoch": 0.00056,
-      "grad_norm": 0.0032236159313470125,
-      "kl": 0.9808710217475891,
+      "grad_norm": 0.0009110283572226763,
+      "kl": 0.4785164334462024,
       "learning_rate": 7.999999851904105e-06,
       "loss": -0.0,
       "step": 56,
-      "step_time": 6.5880885079996006
+      "step_time": 6.473339682004735
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1640,41 +1640,41 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 3.0,
-      "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.375,
-      "completions/mean_terminated_length": 2.375,
+      "completions/max_length": 8.0,
+      "completions/max_terminated_length": 8.0,
+      "completions/mean_length": 2.90625,
+      "completions/mean_terminated_length": 2.90625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.766272783279419,
+      "entropy": 0.7907501719892025,
       "epoch": 0.00057,
-      "frac_reward_zero_std": 1.0,
-      "grad_norm": 5.2334169595269486e-05,
-      "kl": 0.8993725756881759,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0003316142538096756,
+      "kl": 0.13017687457613647,
       "learning_rate": 7.999999836724277e-06,
       "loss": 0.0,
-      "num_tokens": 1341867.0,
-      "reward": 1.3250000476837158,
-      "reward_std": 0.49186936020851135,
+      "num_tokens": 1378379.0,
+      "reward": 2.3249998092651367,
+      "reward_std": 1.4756081104278564,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.375,
-      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_completion_length/mean": 1.9375,
+      "rewards/probe_completion_length/std": 1.1896733045578003,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 1.0,
-      "rewards/probe_shaping_dominance/std": 0.0,
-      "rewards/probe_terminal_raw/mean": 0.0,
-      "rewards/probe_terminal_raw/std": 0.0,
-      "rewards/rollout_reward_func/mean": -1.0,
-      "rewards/rollout_reward_func/std": 0.0,
-      "sampling/importance_sampling_ratio/max": 1.305932879447937,
-      "sampling/importance_sampling_ratio/mean": 0.9440467357635498,
-      "sampling/importance_sampling_ratio/min": 0.6217635273933411,
-      "sampling/sampling_logp_difference/max": 0.3779376745223999,
-      "sampling/sampling_logp_difference/mean": 0.08750458061695099,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.202713131904602,
+      "sampling/importance_sampling_ratio/mean": 0.864936113357544,
+      "sampling/importance_sampling_ratio/min": 0.4689479470252991,
+      "sampling/sampling_logp_difference/max": 0.6789623498916626,
+      "sampling/sampling_logp_difference/mean": 0.0825042799115181,
       "step": 57,
-      "step_time": 9.915378274999057
+      "step_time": 10.657130387999132
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1682,14 +1682,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.7637373208999634,
+      "entropy": 0.7701332122087479,
       "epoch": 0.00058,
-      "grad_norm": 3.501655737636611e-05,
-      "kl": 0.8927765771513805,
+      "grad_norm": 0.00017936780932359397,
+      "kl": 0.12208469014149159,
       "learning_rate": 7.999999820803968e-06,
       "loss": 0.0,
       "step": 58,
-      "step_time": 6.4946643829989625
+      "step_time": 6.01674000300045
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1700,24 +1700,24 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.53125,
-      "completions/mean_terminated_length": 2.53125,
+      "completions/mean_length": 2.71875,
+      "completions/mean_terminated_length": 2.71875,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.6821361035108566,
+      "entropy": 0.5843091569840908,
       "epoch": 0.00059,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 0.00014250715321395546,
-      "kl": 0.5903439885005355,
+      "grad_norm": 0.0009200393105857074,
+      "kl": 0.46439418219961226,
       "learning_rate": 7.99999980414318e-06,
       "loss": 0.0,
-      "num_tokens": 1389227.0,
-      "reward": 1.9812500476837158,
-      "reward_std": 1.2309025526046753,
+      "num_tokens": 1424478.0,
+      "reward": 2.168750047683716,
+      "reward_std": 1.1283552646636963,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.53125,
-      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_completion_length/mean": 1.71875,
+      "rewards/probe_completion_length/std": 0.45680341124534607,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
       "rewards/probe_shaping_dominance/mean": 0.75,
@@ -1726,13 +1726,13 @@
       "rewards/probe_terminal_raw/std": 0.4399413466453552,
       "rewards/rollout_reward_func/mean": -0.5,
       "rewards/rollout_reward_func/std": 0.8798826932907104,
-      "sampling/importance_sampling_ratio/max": 1.239275336265564,
-      "sampling/importance_sampling_ratio/mean": 0.895161509513855,
-      "sampling/importance_sampling_ratio/min": 0.5463025569915771,
-      "sampling/sampling_logp_difference/max": 0.35671401023864746,
-      "sampling/sampling_logp_difference/mean": 0.06356121599674225,
+      "sampling/importance_sampling_ratio/max": 1.380016565322876,
+      "sampling/importance_sampling_ratio/mean": 0.8878653049468994,
+      "sampling/importance_sampling_ratio/min": 0.350737601518631,
+      "sampling/sampling_logp_difference/max": 0.9022061824798584,
+      "sampling/sampling_logp_difference/mean": 0.06674876809120178,
       "step": 59,
-      "step_time": 9.911828085000707
+      "step_time": 10.694612899000276
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1740,14 +1740,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.6835376396775246,
+      "entropy": 0.5868858397006989,
       "epoch": 0.0006,
-      "grad_norm": 0.00015305924171116203,
-      "kl": 0.5899961635004729,
+      "grad_norm": 0.0011626366758719087,
+      "kl": 0.48716597934253514,
       "learning_rate": 7.999999786741913e-06,
       "loss": 0.0,
       "step": 60,
-      "step_time": 6.513966230999358
+      "step_time": 6.897457114002464
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1758,97 +1758,97 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.71875,
-      "completions/mean_terminated_length": 2.71875,
+      "completions/mean_length": 2.5625,
+      "completions/mean_terminated_length": 2.5625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.5099964067339897,
+      "entropy": 0.7314568310976028,
       "epoch": 0.00061,
-      "frac_reward_zero_std": 1.0,
-      "grad_norm": 0.0091713797301054,
-      "kl": 2.064895814655756,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0019117726478725672,
+      "kl": 0.5823338131885976,
       "learning_rate": 7.999999768600167e-06,
-      "loss": 0.0,
-      "num_tokens": 1433796.0,
-      "reward": 2.668750047683716,
-      "reward_std": 1.3496564626693726,
+      "loss": -0.0,
+      "num_tokens": 1475201.0,
+      "reward": 1.8875000476837158,
+      "reward_std": 1.1053389310836792,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.71875,
-      "rewards/probe_completion_length/std": 0.45680341124534607,
+      "rewards/probe_completion_length/mean": 1.5625,
+      "rewards/probe_completion_length/std": 0.504016101360321,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.5,
-      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
-      "rewards/probe_terminal_raw/mean": 0.5,
-      "rewards/probe_terminal_raw/std": 0.5080004930496216,
-      "rewards/rollout_reward_func/mean": 0.0,
-      "rewards/rollout_reward_func/std": 1.0160009860992432,
-      "sampling/importance_sampling_ratio/max": 1.0167537927627563,
-      "sampling/importance_sampling_ratio/mean": 0.7621486783027649,
-      "sampling/importance_sampling_ratio/min": 0.06427010893821716,
-      "sampling/sampling_logp_difference/max": 3.6251673698425293,
-      "sampling/sampling_logp_difference/mean": 0.281258225440979,
+      "rewards/probe_shaping_dominance/mean": 0.8125,
+      "rewards/probe_shaping_dominance/std": 0.3965577781200409,
+      "rewards/probe_terminal_raw/mean": 0.1875,
+      "rewards/probe_terminal_raw/std": 0.3965577781200409,
+      "rewards/rollout_reward_func/mean": -0.625,
+      "rewards/rollout_reward_func/std": 0.7931155562400818,
+      "sampling/importance_sampling_ratio/max": 1.9046052694320679,
+      "sampling/importance_sampling_ratio/mean": 0.9178754091262817,
+      "sampling/importance_sampling_ratio/min": 0.3830260634422302,
+      "sampling/sampling_logp_difference/max": 0.8716261386871338,
+      "sampling/sampling_logp_difference/mean": 0.08674927055835724,
       "step": 61,
-      "step_time": 10.16635897499782
+      "step_time": 10.71165037100036
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_mean": 0.012500000186264515,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.0,
-      "entropy": 0.544323917478323,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.7356374487280846,
       "epoch": 0.00062,
-      "grad_norm": 0.001195584307424724,
-      "kl": 0.8833366855978966,
+      "grad_norm": 0.0006256209453567863,
+      "kl": 0.6181492484174669,
       "learning_rate": 7.99999974971794e-06,
-      "loss": 0.0,
+      "loss": -0.0,
       "step": 62,
-      "step_time": 6.592917243999182
+      "step_time": 6.187538994001443
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 8.0,
-      "completions/max_terminated_length": 8.0,
-      "completions/mean_length": 2.59375,
-      "completions/mean_terminated_length": 2.59375,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.4375,
+      "completions/mean_terminated_length": 2.4375,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.6810576692223549,
+      "entropy": 0.8409040607511997,
       "epoch": 0.00063,
-      "frac_reward_zero_std": 1.0,
-      "grad_norm": 0.00010912255675066262,
-      "kl": 0.03229644891916905,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.01637086272239685,
+      "kl": 3.183713400037959,
       "learning_rate": 7.999999730095235e-06,
-      "loss": 0.0,
-      "num_tokens": 1481625.0,
-      "reward": 2.1062498092651367,
-      "reward_std": 1.7799850702285767,
+      "loss": -0.0001,
+      "num_tokens": 1522612.0,
+      "reward": 1.8250000476837158,
+      "reward_std": 1.2115039825439453,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.65625,
-      "rewards/probe_completion_length/std": 1.4280457496643066,
+      "rewards/probe_completion_length/mean": 1.4375,
+      "rewards/probe_completion_length/std": 0.504016101360321,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.75,
-      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
-      "rewards/probe_terminal_raw/mean": 0.25,
-      "rewards/probe_terminal_raw/std": 0.4399413466453552,
-      "rewards/rollout_reward_func/mean": -0.5,
-      "rewards/rollout_reward_func/std": 0.8798826932907104,
-      "sampling/importance_sampling_ratio/max": 1.4093968868255615,
-      "sampling/importance_sampling_ratio/mean": 0.9041081666946411,
-      "sampling/importance_sampling_ratio/min": 0.6358160972595215,
-      "sampling/sampling_logp_difference/max": 0.4337819218635559,
-      "sampling/sampling_logp_difference/mean": 0.06429628282785416,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.2315932512283325,
+      "sampling/importance_sampling_ratio/mean": 0.8749663829803467,
+      "sampling/importance_sampling_ratio/min": 0.015888746827840805,
+      "sampling/sampling_logp_difference/max": 3.4728269577026367,
+      "sampling/sampling_logp_difference/mean": 0.12395516037940979,
       "step": 63,
-      "step_time": 9.980714073000854
+      "step_time": 10.560535980997884
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1856,14 +1856,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.6828571707010269,
+      "entropy": 0.8305284082889557,
       "epoch": 0.00064,
-      "grad_norm": 8.877379150362685e-05,
-      "kl": 0.03626353240088065,
+      "grad_norm": 0.0032985317520797253,
+      "kl": 0.5883966491091996,
       "learning_rate": 7.99999970973205e-06,
-      "loss": 0.0,
+      "loss": -0.0001,
       "step": 64,
-      "step_time": 6.516721850999602
+      "step_time": 6.416247896999266
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1874,39 +1874,39 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.40625,
-      "completions/mean_terminated_length": 2.40625,
+      "completions/mean_length": 2.5,
+      "completions/mean_terminated_length": 2.5,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.8135795295238495,
+      "entropy": 0.7524066865444183,
       "epoch": 0.00065,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 6.834424857515842e-05,
-      "kl": 0.032632274171191966,
+      "grad_norm": 0.00010499630297999829,
+      "kl": 0.20387221802957356,
       "learning_rate": 7.999999688628386e-06,
       "loss": 0.0,
-      "num_tokens": 1527985.0,
-      "reward": 1.8562500476837158,
-      "reward_std": 1.2790968418121338,
+      "num_tokens": 1571590.0,
+      "reward": 1.4812500476837158,
+      "reward_std": 0.507007360458374,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.40625,
-      "rewards/probe_completion_length/std": 0.49899089336395264,
-      "rewards/probe_invalid_count/mean": 0.0,
-      "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.75,
-      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
-      "rewards/probe_terminal_raw/mean": 0.25,
-      "rewards/probe_terminal_raw/std": 0.4399413466453552,
-      "rewards/rollout_reward_func/mean": -0.5,
-      "rewards/rollout_reward_func/std": 0.8798826932907104,
-      "sampling/importance_sampling_ratio/max": 1.3485047817230225,
-      "sampling/importance_sampling_ratio/mean": 0.9244140386581421,
-      "sampling/importance_sampling_ratio/min": 0.4989074468612671,
-      "sampling/sampling_logp_difference/max": 0.4888077974319458,
-      "sampling/sampling_logp_difference/mean": 0.09029445052146912,
+      "rewards/probe_completion_length/mean": 1.5,
+      "rewards/probe_completion_length/std": 0.5080004930496216,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.4499962329864502,
+      "sampling/importance_sampling_ratio/mean": 0.888777494430542,
+      "sampling/importance_sampling_ratio/min": 0.5364581942558289,
+      "sampling/sampling_logp_difference/max": 0.5399905443191528,
+      "sampling/sampling_logp_difference/mean": 0.07261274755001068,
       "step": 65,
-      "step_time": 10.491214556999694
+      "step_time": 10.264871265002512
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1914,14 +1914,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.8274623081088066,
+      "entropy": 0.759266197681427,
       "epoch": 0.00066,
-      "grad_norm": 8.216574497055262e-05,
-      "kl": 0.03480283582757693,
+      "grad_norm": 0.00010316984116798267,
+      "kl": 0.20505999261513352,
       "learning_rate": 7.999999666784243e-06,
       "loss": 0.0,
       "step": 66,
-      "step_time": 6.049761105999096
+      "step_time": 5.990639261999604
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1930,41 +1930,41 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 3.0,
-      "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.5,
-      "completions/mean_terminated_length": 2.5,
+      "completions/max_length": 7.0,
+      "completions/max_terminated_length": 7.0,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.8280301168560982,
+      "entropy": 0.8478573858737946,
       "epoch": 0.00067,
       "frac_reward_zero_std": 1.0,
-      "grad_norm": 8.851613529259339e-05,
-      "kl": 0.19343006524650264,
+      "grad_norm": 0.00017549873155076057,
+      "kl": 0.09552802673715632,
       "learning_rate": 7.999999644199619e-06,
       "loss": 0.0,
-      "num_tokens": 1575885.0,
-      "reward": 1.4500000476837158,
-      "reward_std": 0.5080004930496216,
+      "num_tokens": 1621069.0,
+      "reward": 1.7000000476837158,
+      "reward_std": 1.2443419694900513,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.5,
-      "rewards/probe_completion_length/std": 0.5080004930496216,
-      "rewards/probe_invalid_count/mean": 0.0,
-      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.71875,
+      "rewards/probe_completion_length/std": 1.2504031658172607,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
       "rewards/probe_shaping_dominance/mean": 1.0,
       "rewards/probe_shaping_dominance/std": 0.0,
       "rewards/probe_terminal_raw/mean": 0.0,
       "rewards/probe_terminal_raw/std": 0.0,
       "rewards/rollout_reward_func/mean": -1.0,
       "rewards/rollout_reward_func/std": 0.0,
-      "sampling/importance_sampling_ratio/max": 1.5432251691818237,
-      "sampling/importance_sampling_ratio/mean": 0.9077194929122925,
-      "sampling/importance_sampling_ratio/min": 0.4163496196269989,
-      "sampling/sampling_logp_difference/max": 0.6807805299758911,
-      "sampling/sampling_logp_difference/mean": 0.09229114651679993,
+      "sampling/importance_sampling_ratio/max": 1.1542682647705078,
+      "sampling/importance_sampling_ratio/mean": 0.8189550638198853,
+      "sampling/importance_sampling_ratio/min": 0.36865320801734924,
+      "sampling/sampling_logp_difference/max": 0.900229811668396,
+      "sampling/sampling_logp_difference/mean": 0.10676399618387222,
       "step": 67,
-      "step_time": 10.033269270999881
+      "step_time": 11.213862350001364
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -1972,57 +1972,57 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.8344325572252274,
+      "entropy": 0.8443719670176506,
       "epoch": 0.00068,
-      "grad_norm": 6.554734864039347e-05,
-      "kl": 0.19161523527145619,
+      "grad_norm": 0.00021201952768024057,
+      "kl": 0.10493764441343956,
       "learning_rate": 7.999999620874517e-06,
       "loss": 0.0,
       "step": 68,
-      "step_time": 7.017202997999448
+      "step_time": 6.09188792200257
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_mean": 0.035416667349636555,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.0,
+      "clip_ratio/region_mean": 0.035416667349636555,
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.625,
-      "completions/mean_terminated_length": 2.625,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.5219790041446686,
+      "entropy": 0.8209497034549713,
       "epoch": 0.00069,
-      "frac_reward_zero_std": 1.0,
-      "grad_norm": 0.000104649989225436,
-      "kl": 1.053569495677948,
+      "frac_reward_zero_std": 0.5,
+      "grad_norm": 0.0020198614802211523,
+      "kl": 0.8513226807117462,
       "learning_rate": 7.999999596808934e-06,
-      "loss": 0.0,
-      "num_tokens": 1622330.0,
-      "reward": 2.575000047683716,
-      "reward_std": 1.4312187433242798,
+      "loss": -0.0001,
+      "num_tokens": 1667152.0,
+      "reward": 2.481250047683716,
+      "reward_std": 1.3674646615982056,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.625,
-      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_completion_length/mean": 1.65625,
+      "rewards/probe_completion_length/std": 0.4825586974620819,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.5,
-      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
-      "rewards/probe_terminal_raw/mean": 0.5,
-      "rewards/probe_terminal_raw/std": 0.5080004930496216,
-      "rewards/rollout_reward_func/mean": 0.0,
-      "rewards/rollout_reward_func/std": 1.0160009860992432,
-      "sampling/importance_sampling_ratio/max": 1.0400450229644775,
-      "sampling/importance_sampling_ratio/mean": 0.9155407547950745,
-      "sampling/importance_sampling_ratio/min": 0.7058902382850647,
-      "sampling/sampling_logp_difference/max": 0.24529099464416504,
-      "sampling/sampling_logp_difference/mean": 0.04126736894249916,
+      "rewards/probe_shaping_dominance/mean": 0.5625,
+      "rewards/probe_shaping_dominance/std": 0.504016101360321,
+      "rewards/probe_terminal_raw/mean": 0.4375,
+      "rewards/probe_terminal_raw/std": 0.504016101360321,
+      "rewards/rollout_reward_func/mean": -0.125,
+      "rewards/rollout_reward_func/std": 1.008032202720642,
+      "sampling/importance_sampling_ratio/max": 1.8785016536712646,
+      "sampling/importance_sampling_ratio/mean": 0.9165289402008057,
+      "sampling/importance_sampling_ratio/min": 0.4227677583694458,
+      "sampling/sampling_logp_difference/max": 0.8007502555847168,
+      "sampling/sampling_logp_difference/mean": 0.10074363648891449,
       "step": 69,
-      "step_time": 9.983138362999853
+      "step_time": 10.905393431999983
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -2030,14 +2030,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.5127745680510998,
+      "entropy": 0.8409527391195297,
       "epoch": 0.0007,
-      "grad_norm": 0.00018927347264252603,
-      "kl": 1.0697944089770317,
+      "grad_norm": 0.009374535642564297,
+      "kl": 0.7900811154395342,
       "learning_rate": 7.999999572002872e-06,
-      "loss": 0.0,
+      "loss": -0.0001,
       "step": 70,
-      "step_time": 6.019172412000444
+      "step_time": 6.190164835001269
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -2048,97 +2048,155 @@
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.8125,
-      "completions/mean_terminated_length": 2.8125,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.607807818800211,
+      "entropy": 0.640807468444109,
       "epoch": 0.00071,
-      "frac_reward_zero_std": 1.0,
-      "grad_norm": 0.0003802482970058918,
-      "kl": 0.21827068779384717,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0027836360968649387,
+      "kl": 0.6423672240234737,
       "learning_rate": 7.999999546456332e-06,
-      "loss": 0.0,
-      "num_tokens": 1671382.0,
-      "reward": 2.262500047683716,
-      "reward_std": 1.0606601238250732,
+      "loss": -0.0001,
+      "num_tokens": 1713194.0,
+      "reward": 1.9500000476837158,
+      "reward_std": 1.0776318311691284,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.8125,
-      "rewards/probe_completion_length/std": 0.3965577781200409,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.75,
-      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
-      "rewards/probe_terminal_raw/mean": 0.25,
-      "rewards/probe_terminal_raw/std": 0.4399413466453552,
-      "rewards/rollout_reward_func/mean": -0.5,
-      "rewards/rollout_reward_func/std": 0.8798826932907104,
-      "sampling/importance_sampling_ratio/max": 2.478405714035034,
-      "sampling/importance_sampling_ratio/mean": 0.9307346343994141,
-      "sampling/importance_sampling_ratio/min": 0.17532044649124146,
-      "sampling/sampling_logp_difference/max": 1.585439682006836,
-      "sampling/sampling_logp_difference/mean": 0.10700278729200363,
+      "rewards/probe_shaping_dominance/mean": 0.8125,
+      "rewards/probe_shaping_dominance/std": 0.3965577781200409,
+      "rewards/probe_terminal_raw/mean": 0.1875,
+      "rewards/probe_terminal_raw/std": 0.3965577781200409,
+      "rewards/rollout_reward_func/mean": -0.625,
+      "rewards/rollout_reward_func/std": 0.7931155562400818,
+      "sampling/importance_sampling_ratio/max": 1.3413078784942627,
+      "sampling/importance_sampling_ratio/mean": 0.8879814147949219,
+      "sampling/importance_sampling_ratio/min": 0.3190351724624634,
+      "sampling/sampling_logp_difference/max": 1.076015830039978,
+      "sampling/sampling_logp_difference/mean": 0.08176332712173462,
       "step": 71,
-      "step_time": 10.181540167999628
+      "step_time": 10.522709103001034
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_mean": 0.012500000186264515,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.0,
-      "entropy": 0.6031206957995892,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.6309941411018372,
       "epoch": 0.00072,
-      "grad_norm": 0.00040189133142121136,
-      "kl": 0.2209149764967151,
+      "grad_norm": 0.002217747736722231,
+      "kl": 1.143750467152131,
       "learning_rate": 7.999999520169313e-06,
-      "loss": 0.0,
+      "loss": -0.0001,
       "step": 72,
-      "step_time": 7.0209326659978615
+      "step_time": 5.988715680998212
     },
     {
       "clip_ratio/high_max": 0.0,
       "clip_ratio/high_mean": 0.0,
-      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
       "clip_ratio/low_min": 0.0,
-      "clip_ratio/region_mean": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
       "completions/clipped_ratio": 0.0,
       "completions/max_length": 3.0,
       "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.4375,
-      "completions/mean_terminated_length": 2.4375,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.7966198176145554,
+      "entropy": 0.6769370436668396,
       "epoch": 0.00073,
-      "frac_reward_zero_std": 1.0,
-      "grad_norm": 5.5945911299204454e-05,
-      "kl": 0.019231000915169716,
+      "frac_reward_zero_std": 0.5,
+      "grad_norm": 0.011145965196192265,
+      "kl": 0.6622413620352745,
       "learning_rate": 7.999999493141815e-06,
+      "loss": -0.0002,
+      "num_tokens": 1760736.0,
+      "reward": 2.387500047683716,
+      "reward_std": 1.1341474056243896,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.78125,
+      "rewards/probe_completion_length/std": 0.420013427734375,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.6875,
+      "rewards/probe_shaping_dominance/std": 0.4709290862083435,
+      "rewards/probe_terminal_raw/mean": 0.3125,
+      "rewards/probe_terminal_raw/std": 0.4709290862083435,
+      "rewards/rollout_reward_func/mean": -0.375,
+      "rewards/rollout_reward_func/std": 0.941858172416687,
+      "sampling/importance_sampling_ratio/max": 2.0832128524780273,
+      "sampling/importance_sampling_ratio/mean": 0.8911354541778564,
+      "sampling/importance_sampling_ratio/min": 0.2129342257976532,
+      "sampling/sampling_logp_difference/max": 1.4288380146026611,
+      "sampling/sampling_logp_difference/mean": 0.13255563378334045,
+      "step": 73,
+      "step_time": 10.575429333001011
+    },
+    {
+      "clip_ratio/high_max": 0.02083333395421505,
+      "clip_ratio/high_mean": 0.010416666977107525,
+      "clip_ratio/low_mean": 0.06458333507180214,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.07500000111758709,
+      "entropy": 0.642921295017004,
+      "epoch": 0.00074,
+      "grad_norm": 0.0028342357836663723,
+      "kl": 1.2112234272062778,
+      "learning_rate": 7.999999465373833e-06,
+      "loss": -0.0002,
+      "step": 74,
+      "step_time": 5.997502170999724
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 3.0,
+      "completions/mean_terminated_length": 3.0,
+      "completions/min_length": 3.0,
+      "completions/min_terminated_length": 3.0,
+      "entropy": 0.4875616766512394,
+      "epoch": 0.00075,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00022702451678924263,
+      "kl": 0.5974396051169606,
+      "learning_rate": 7.999999436865376e-06,
       "loss": 0.0,
-      "num_tokens": 1721707.0,
-      "reward": 1.3875000476837158,
-      "reward_std": 0.504016101360321,
+      "num_tokens": 1806616.0,
+      "reward": 2.450000047683716,
+      "reward_std": 0.8798826932907104,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.4375,
-      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_completion_length/mean": 2.0,
+      "rewards/probe_completion_length/std": 0.0,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 1.0,
-      "rewards/probe_shaping_dominance/std": 0.0,
-      "rewards/probe_terminal_raw/mean": 0.0,
-      "rewards/probe_terminal_raw/std": 0.0,
-      "rewards/rollout_reward_func/mean": -1.0,
-      "rewards/rollout_reward_func/std": 0.0,
-      "sampling/importance_sampling_ratio/max": 1.1157020330429077,
-      "sampling/importance_sampling_ratio/mean": 0.8640221953392029,
-      "sampling/importance_sampling_ratio/min": 0.5155755877494812,
-      "sampling/sampling_logp_difference/max": 0.4500473737716675,
-      "sampling/sampling_logp_difference/mean": 0.07616540789604187,
-      "step": 73,
-      "step_time": 10.120440676000726
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 2.3010740280151367,
+      "sampling/importance_sampling_ratio/mean": 1.0142158269882202,
+      "sampling/importance_sampling_ratio/min": 0.37630370259284973,
+      "sampling/sampling_logp_difference/max": 1.117997646331787,
+      "sampling/sampling_logp_difference/mean": 0.10312129557132721,
+      "step": 75,
+      "step_time": 10.39726536499984
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -2146,14 +2204,14 @@
       "clip_ratio/low_mean": 0.0,
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
-      "entropy": 0.7941807508468628,
-      "epoch": 0.00074,
-      "grad_norm": 2.562938243499957e-05,
-      "kl": 0.009964701312128454,
-      "learning_rate": 7.999999465373833e-06,
+      "entropy": 0.48410840332508087,
+      "epoch": 0.00076,
+      "grad_norm": 0.0004807643999811262,
+      "kl": 0.6427391922770767,
+      "learning_rate": 7.999999407616439e-06,
       "loss": 0.0,
-      "step": 74,
-      "step_time": 6.081612156000119
+      "step": 76,
+      "step_time": 6.295117113002561
     },
     {
       "clip_ratio/high_max": 0.0,
@@ -2162,46 +2220,6499 @@
       "clip_ratio/low_min": 0.0,
       "clip_ratio/region_mean": 0.0,
       "completions/clipped_ratio": 0.0,
-      "completions/max_length": 3.0,
-      "completions/max_terminated_length": 3.0,
-      "completions/mean_length": 2.65625,
-      "completions/mean_terminated_length": 2.65625,
+      "completions/max_length": 7.0,
+      "completions/max_terminated_length": 7.0,
+      "completions/mean_length": 2.84375,
+      "completions/mean_terminated_length": 2.84375,
       "completions/min_length": 2.0,
       "completions/min_terminated_length": 2.0,
-      "entropy": 0.5862512141466141,
-      "epoch": 0.00075,
+      "entropy": 0.7104897350072861,
+      "epoch": 0.00077,
       "frac_reward_zero_std": 0.75,
-      "grad_norm": 0.0030402804259210825,
-      "kl": 2.1613324359059334,
-      "learning_rate": 7.999999436865376e-06,
+      "grad_norm": 0.0019423841731622815,
+      "kl": 0.7007742009591311,
+      "learning_rate": 7.999999377627022e-06,
       "loss": -0.0001,
-      "num_tokens": 1769976.0,
-      "reward": 1.9812500476837158,
-      "reward_std": 1.0620847940444946,
+      "num_tokens": 1854875.0,
+      "reward": 2.0437498092651367,
+      "reward_std": 1.3526406288146973,
       "rewards/format_guard/mean": -0.05000000074505806,
       "rewards/format_guard/std": 0.0,
-      "rewards/probe_completion_length/mean": 1.65625,
-      "rewards/probe_completion_length/std": 0.4825586974620819,
+      "rewards/probe_completion_length/mean": 1.90625,
+      "rewards/probe_completion_length/std": 1.201058030128479,
       "rewards/probe_invalid_count/mean": 0.0,
       "rewards/probe_invalid_count/std": 0.0,
-      "rewards/probe_shaping_dominance/mean": 0.8125,
-      "rewards/probe_shaping_dominance/std": 0.3965577781200409,
+      "rewards/probe_shaping_dominance/mean": 0.90625,
+      "rewards/probe_shaping_dominance/std": 0.2961445748806,
+      "rewards/probe_terminal_raw/mean": 0.09375,
+      "rewards/probe_terminal_raw/std": 0.2961445748806,
+      "rewards/rollout_reward_func/mean": -0.8125,
+      "rewards/rollout_reward_func/std": 0.5922891497612,
+      "sampling/importance_sampling_ratio/max": 1.0384966135025024,
+      "sampling/importance_sampling_ratio/mean": 0.7873358130455017,
+      "sampling/importance_sampling_ratio/min": 0.16774220764636993,
+      "sampling/sampling_logp_difference/max": 1.7101855278015137,
+      "sampling/sampling_logp_difference/mean": 0.12416968494653702,
+      "step": 77,
+      "step_time": 10.378455145999396
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.031250000931322575,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.031250000931322575,
+      "entropy": 0.712644524872303,
+      "epoch": 0.00078,
+      "grad_norm": 0.001622700714506209,
+      "kl": 0.9561241322662681,
+      "learning_rate": 7.999999346897126e-06,
+      "loss": -0.0001,
+      "step": 78,
+      "step_time": 6.033231712997804
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.4954247400164604,
+      "epoch": 0.00079,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 7.158728112699464e-05,
+      "kl": 1.1242153272032738,
+      "learning_rate": 7.99999931542675e-06,
+      "loss": 0.0,
+      "num_tokens": 1901828.0,
+      "reward": 2.731250047683716,
+      "reward_std": 1.2885193824768066,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.78125,
+      "rewards/probe_completion_length/std": 0.420013427734375,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5,
+      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
+      "rewards/probe_terminal_raw/mean": 0.5,
+      "rewards/probe_terminal_raw/std": 0.5080004930496216,
+      "rewards/rollout_reward_func/mean": 0.0,
+      "rewards/rollout_reward_func/std": 1.0160009860992432,
+      "sampling/importance_sampling_ratio/max": 1.4480345249176025,
+      "sampling/importance_sampling_ratio/mean": 0.9386520385742188,
+      "sampling/importance_sampling_ratio/min": 0.6387155055999756,
+      "sampling/sampling_logp_difference/max": 0.4249168634414673,
+      "sampling/sampling_logp_difference/mean": 0.05157982185482979,
+      "step": 79,
+      "step_time": 11.405814464003925
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.5006568469107151,
+      "epoch": 0.0008,
+      "grad_norm": 0.0001417232269886881,
+      "kl": 1.1419733166694641,
+      "learning_rate": 7.999999283215897e-06,
+      "loss": 0.0,
+      "step": 80,
+      "step_time": 6.161420755997824
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.59375,
+      "completions/mean_terminated_length": 2.59375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6742436736822128,
+      "epoch": 0.00081,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.00282097514718771,
+      "kl": 1.1596152225974947,
+      "learning_rate": 7.999999250264562e-06,
+      "loss": -0.0001,
+      "num_tokens": 1953232.0,
+      "reward": 1.9187500476837158,
+      "reward_std": 1.092034935951233,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.59375,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.8125,
+      "rewards/probe_shaping_dominance/std": 0.3965577781200409,
       "rewards/probe_terminal_raw/mean": 0.1875,
       "rewards/probe_terminal_raw/std": 0.3965577781200409,
       "rewards/rollout_reward_func/mean": -0.625,
       "rewards/rollout_reward_func/std": 0.7931155562400818,
-      "sampling/importance_sampling_ratio/max": 1.468144416809082,
-      "sampling/importance_sampling_ratio/mean": 0.9493575096130371,
-      "sampling/importance_sampling_ratio/min": 0.03764805570244789,
-      "sampling/sampling_logp_difference/max": 3.1981916427612305,
-      "sampling/sampling_logp_difference/mean": 0.11449257284402847,
-      "step": 75,
-      "step_time": 10.459286816998429
+      "sampling/importance_sampling_ratio/max": 1.3345328569412231,
+      "sampling/importance_sampling_ratio/mean": 0.9116201400756836,
+      "sampling/importance_sampling_ratio/min": 0.0795779824256897,
+      "sampling/sampling_logp_difference/max": 2.446382999420166,
+      "sampling/sampling_logp_difference/mean": 0.12286892533302307,
+      "step": 81,
+      "step_time": 10.233006808000937
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "entropy": 0.6922024786472321,
+      "epoch": 0.00082,
+      "grad_norm": 0.001407204894348979,
+      "kl": 0.876353993313387,
+      "learning_rate": 7.999999216572749e-06,
+      "loss": -0.0001,
+      "step": 82,
+      "step_time": 6.428627160998076
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7211672216653824,
+      "epoch": 0.00083,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00017314944125246257,
+      "kl": 0.26839208393357694,
+      "learning_rate": 7.999999182140456e-06,
+      "loss": 0.0,
+      "num_tokens": 2001699.0,
+      "reward": 1.5750000476837158,
+      "reward_std": 0.49186936020851135,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 2.0702590942382812,
+      "sampling/importance_sampling_ratio/mean": 0.918670654296875,
+      "sampling/importance_sampling_ratio/min": 0.31061917543411255,
+      "sampling/sampling_logp_difference/max": 0.9575868248939514,
+      "sampling/sampling_logp_difference/mean": 0.09307387471199036,
+      "step": 83,
+      "step_time": 10.531592241997714
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7358985021710396,
+      "epoch": 0.00084,
+      "grad_norm": 0.00017405653488822281,
+      "kl": 0.2663446672777354,
+      "learning_rate": 7.999999146967684e-06,
+      "loss": 0.0,
+      "step": 84,
+      "step_time": 6.026314235999962
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.59375,
+      "completions/mean_terminated_length": 2.59375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8648177459836006,
+      "epoch": 0.00085,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.000491270562633872,
+      "kl": 0.09633485117228702,
+      "learning_rate": 7.999999111054434e-06,
+      "loss": 0.0,
+      "num_tokens": 2049769.0,
+      "reward": 1.5437500476837158,
+      "reward_std": 0.498990923166275,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.59375,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 0.9729283452033997,
+      "sampling/importance_sampling_ratio/mean": 0.7784683704376221,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 1.6030910015106201,
+      "sampling/sampling_logp_difference/mean": 0.12049049139022827,
+      "step": 85,
+      "step_time": 10.301046949998636
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8687215521931648,
+      "epoch": 0.00086,
+      "grad_norm": 0.0005266318912617862,
+      "kl": 0.09560301806777716,
+      "learning_rate": 7.999999074400703e-06,
+      "loss": 0.0,
+      "step": 86,
+      "step_time": 6.827429698998458
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6539479978382587,
+      "epoch": 0.00087,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.415531242964789e-05,
+      "kl": 0.36288101226091385,
+      "learning_rate": 7.999999037006494e-06,
+      "loss": 0.0,
+      "num_tokens": 2096587.0,
+      "reward": 1.6062500476837158,
+      "reward_std": 0.4825586974620819,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.65625,
+      "rewards/probe_completion_length/std": 0.4825586974620819,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.364053726196289,
+      "sampling/importance_sampling_ratio/mean": 0.9139212369918823,
+      "sampling/importance_sampling_ratio/min": 0.46633586287498474,
+      "sampling/sampling_logp_difference/max": 0.7080342769622803,
+      "sampling/sampling_logp_difference/mean": 0.06971700489521027,
+      "step": 87,
+      "step_time": 10.411047061999852
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6638137400150299,
+      "epoch": 0.00088,
+      "grad_norm": 6.288591248448938e-05,
+      "kl": 0.36502306535840034,
+      "learning_rate": 7.999998998871805e-06,
+      "loss": 0.0,
+      "step": 88,
+      "step_time": 6.455294155004594
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6723155006766319,
+      "epoch": 0.00089,
+      "frac_reward_zero_std": 0.5,
+      "grad_norm": 0.004749135114252567,
+      "kl": 2.617519076447934,
+      "learning_rate": 7.999998959996637e-06,
+      "loss": -0.0002,
+      "num_tokens": 2144825.0,
+      "reward": 2.262500047683716,
+      "reward_std": 1.0606601238250732,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.78125,
+      "rewards/probe_completion_length/std": 0.420013427734375,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 2.2531893253326416,
+      "sampling/importance_sampling_ratio/mean": 0.9639551639556885,
+      "sampling/importance_sampling_ratio/min": 0.1937435120344162,
+      "sampling/sampling_logp_difference/max": 1.603487491607666,
+      "sampling/sampling_logp_difference/mean": 0.10713129490613937,
+      "step": 89,
+      "step_time": 10.420832339999833
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.656385749578476,
+      "epoch": 0.0009,
+      "grad_norm": 0.0035466940607875586,
+      "kl": 2.387696197256446,
+      "learning_rate": 7.99999892038099e-06,
+      "loss": -0.0002,
+      "step": 90,
+      "step_time": 6.442203360002168
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.53125,
+      "completions/mean_terminated_length": 2.53125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8181677013635635,
+      "epoch": 0.00091,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0019888803362846375,
+      "kl": 0.5702489176765084,
+      "learning_rate": 7.999998880024863e-06,
+      "loss": -0.0001,
+      "num_tokens": 2192809.0,
+      "reward": 1.9187500476837158,
+      "reward_std": 1.1773227453231812,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.53125,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.1860374212265015,
+      "sampling/importance_sampling_ratio/mean": 0.8859227895736694,
+      "sampling/importance_sampling_ratio/min": 0.1813456267118454,
+      "sampling/sampling_logp_difference/max": 1.622064471244812,
+      "sampling/sampling_logp_difference/mean": 0.09506501257419586,
+      "step": 91,
+      "step_time": 10.108775123004307
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.800042949616909,
+      "epoch": 0.00092,
+      "grad_norm": 0.001467581489123404,
+      "kl": 0.5681175279896706,
+      "learning_rate": 7.999998838928257e-06,
+      "loss": -0.0001,
+      "step": 92,
+      "step_time": 6.406815487996937
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.40625,
+      "completions/mean_terminated_length": 2.40625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.716269675642252,
+      "epoch": 0.00093,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.907718488946557e-05,
+      "kl": 0.024005705378840503,
+      "learning_rate": 7.999998797091172e-06,
+      "loss": 0.0,
+      "num_tokens": 2241682.0,
+      "reward": 1.8562500476837158,
+      "reward_std": 1.2790968418121338,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.40625,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.3916001319885254,
+      "sampling/importance_sampling_ratio/mean": 0.9141188859939575,
+      "sampling/importance_sampling_ratio/min": 0.6723463535308838,
+      "sampling/sampling_logp_difference/max": 0.3769274950027466,
+      "sampling/sampling_logp_difference/mean": 0.05592293664813042,
+      "step": 93,
+      "step_time": 10.462285012999928
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6890156902372837,
+      "epoch": 0.00094,
+      "grad_norm": 0.00013785925693809986,
+      "kl": 0.031552536704111844,
+      "learning_rate": 7.999998754513608e-06,
+      "loss": 0.0,
+      "step": 94,
+      "step_time": 5.955127831999562
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.59375,
+      "completions/mean_terminated_length": 2.59375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.5587945282459259,
+      "epoch": 0.00095,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00024127251526806504,
+      "kl": 0.9856726005673409,
+      "learning_rate": 7.999998711195565e-06,
+      "loss": 0.0,
+      "num_tokens": 2286451.0,
+      "reward": 2.543750047683716,
+      "reward_std": 1.4560080766677856,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.59375,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5,
+      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
+      "rewards/probe_terminal_raw/mean": 0.5,
+      "rewards/probe_terminal_raw/std": 0.5080004930496216,
+      "rewards/rollout_reward_func/mean": 0.0,
+      "rewards/rollout_reward_func/std": 1.0160009860992432,
+      "sampling/importance_sampling_ratio/max": 1.0284326076507568,
+      "sampling/importance_sampling_ratio/mean": 0.8956028819084167,
+      "sampling/importance_sampling_ratio/min": 0.4129232168197632,
+      "sampling/sampling_logp_difference/max": 0.7938406467437744,
+      "sampling/sampling_logp_difference/mean": 0.06116846576333046,
+      "step": 95,
+      "step_time": 10.090581595999538
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.5542779862880707,
+      "epoch": 0.00096,
+      "grad_norm": 0.0001853352296166122,
+      "kl": 0.9771002680063248,
+      "learning_rate": 7.999998667137043e-06,
+      "loss": 0.0,
+      "step": 96,
+      "step_time": 6.352050979998239
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.875,
+      "completions/mean_terminated_length": 2.875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.4173930324614048,
+      "epoch": 0.00097,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.00037545509985648096,
+      "kl": 0.9331283923238516,
+      "learning_rate": 7.999998622338041e-06,
+      "loss": 0.0,
+      "num_tokens": 2336439.0,
+      "reward": 2.762500047683716,
+      "reward_std": 1.1760376691818237,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.875,
+      "rewards/probe_completion_length/std": 0.33601075410842896,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.53125,
+      "rewards/probe_shaping_dominance/std": 0.507007360458374,
+      "rewards/probe_terminal_raw/mean": 0.46875,
+      "rewards/probe_terminal_raw/std": 0.507007360458374,
+      "rewards/rollout_reward_func/mean": -0.0625,
+      "rewards/rollout_reward_func/std": 1.014014720916748,
+      "sampling/importance_sampling_ratio/max": 1.4940528869628906,
+      "sampling/importance_sampling_ratio/mean": 0.9181157350540161,
+      "sampling/importance_sampling_ratio/min": 0.556973397731781,
+      "sampling/sampling_logp_difference/max": 0.5000306367874146,
+      "sampling/sampling_logp_difference/mean": 0.05192491039633751,
+      "step": 97,
+      "step_time": 10.801652450001711
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.4088115468621254,
+      "epoch": 0.00098,
+      "grad_norm": 0.0004249197954777628,
+      "kl": 0.9439695589244366,
+      "learning_rate": 7.999998576798562e-06,
+      "loss": 0.0,
+      "step": 98,
+      "step_time": 6.155849446995489
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.84375,
+      "completions/mean_terminated_length": 2.84375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.5160295888781548,
+      "epoch": 0.00099,
+      "frac_reward_zero_std": 0.5,
+      "grad_norm": 0.0072716595605015755,
+      "kl": 4.22201424402374,
+      "learning_rate": 7.999998530518601e-06,
+      "loss": -0.0,
+      "num_tokens": 2379722.0,
+      "reward": 2.418750047683716,
+      "reward_std": 1.106706142425537,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.84375,
+      "rewards/probe_completion_length/std": 0.3689020276069641,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.6875,
+      "rewards/probe_shaping_dominance/std": 0.4709290862083435,
+      "rewards/probe_terminal_raw/mean": 0.3125,
+      "rewards/probe_terminal_raw/std": 0.4709290862083435,
+      "rewards/rollout_reward_func/mean": -0.375,
+      "rewards/rollout_reward_func/std": 0.941858172416687,
+      "sampling/importance_sampling_ratio/max": 1.3872947692871094,
+      "sampling/importance_sampling_ratio/mean": 0.8936517238616943,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 1.780822515487671,
+      "sampling/sampling_logp_difference/mean": 0.09026888012886047,
+      "step": 99,
+      "step_time": 10.53207572199608
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.5115218348801136,
+      "epoch": 0.001,
+      "grad_norm": 0.0031906350050121546,
+      "kl": 2.792147676460445,
+      "learning_rate": 7.999998483498162e-06,
+      "loss": -0.0,
+      "step": 100,
+      "step_time": 5.868994353000744
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5,
+      "completions/mean_terminated_length": 2.5,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8312729708850384,
+      "epoch": 0.00101,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00014353707956615835,
+      "kl": 0.6586928470060229,
+      "learning_rate": 7.999998435737244e-06,
+      "loss": 0.0,
+      "num_tokens": 2428865.0,
+      "reward": 1.9500000476837158,
+      "reward_std": 1.2443419694900513,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5,
+      "rewards/probe_completion_length/std": 0.5080004930496216,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.398367166519165,
+      "sampling/importance_sampling_ratio/mean": 0.9189577698707581,
+      "sampling/importance_sampling_ratio/min": 0.6098452806472778,
+      "sampling/sampling_logp_difference/max": 0.5889620780944824,
+      "sampling/sampling_logp_difference/mean": 0.07546409219503403,
+      "step": 101,
+      "step_time": 10.53599652000048
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8421025648713112,
+      "epoch": 0.00102,
+      "grad_norm": 0.00014895916683599353,
+      "kl": 0.6579505642876029,
+      "learning_rate": 7.999998387235846e-06,
+      "loss": 0.0,
+      "step": 102,
+      "step_time": 6.442589998996482
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 7.0,
+      "completions/max_terminated_length": 7.0,
+      "completions/mean_length": 3.0,
+      "completions/mean_terminated_length": 3.0,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.4542905166745186,
+      "epoch": 0.00103,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00018467936024535447,
+      "kl": 1.264952726662159,
+      "learning_rate": 7.99999833799397e-06,
+      "loss": 0.0,
+      "num_tokens": 2476084.0,
+      "reward": 2.9812498092651367,
+      "reward_std": 1.3792091608047485,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 2.03125,
+      "rewards/probe_completion_length/std": 0.9666828513145447,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5,
+      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
+      "rewards/probe_terminal_raw/mean": 0.5,
+      "rewards/probe_terminal_raw/std": 0.5080004930496216,
+      "rewards/rollout_reward_func/mean": 0.0,
+      "rewards/rollout_reward_func/std": 1.0160009860992432,
+      "sampling/importance_sampling_ratio/max": 1.6235320568084717,
+      "sampling/importance_sampling_ratio/mean": 0.9240143299102783,
+      "sampling/importance_sampling_ratio/min": 0.3445168435573578,
+      "sampling/sampling_logp_difference/max": 0.7978925704956055,
+      "sampling/sampling_logp_difference/mean": 0.06782019138336182,
+      "step": 103,
+      "step_time": 10.232398630996613
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.4663286581635475,
+      "epoch": 0.00104,
+      "grad_norm": 0.0002121606084983796,
+      "kl": 1.2690180391073227,
+      "learning_rate": 7.999998288011616e-06,
+      "loss": 0.0,
+      "step": 104,
+      "step_time": 6.369392825999967
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.59375,
+      "completions/mean_terminated_length": 2.59375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.5813856143504381,
+      "epoch": 0.00105,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.004445977043360472,
+      "kl": 0.5892454935237765,
+      "learning_rate": 7.999998237288781e-06,
+      "loss": -0.0,
+      "num_tokens": 2520485.0,
+      "reward": 2.481250047683716,
+      "reward_std": 1.43649160861969,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.59375,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.53125,
+      "rewards/probe_shaping_dominance/std": 0.507007360458374,
+      "rewards/probe_terminal_raw/mean": 0.46875,
+      "rewards/probe_terminal_raw/std": 0.507007360458374,
+      "rewards/rollout_reward_func/mean": -0.0625,
+      "rewards/rollout_reward_func/std": 1.014014720916748,
+      "sampling/importance_sampling_ratio/max": 1.263745665550232,
+      "sampling/importance_sampling_ratio/mean": 0.915697455406189,
+      "sampling/importance_sampling_ratio/min": 0.44258958101272583,
+      "sampling/sampling_logp_difference/max": 0.7512234449386597,
+      "sampling/sampling_logp_difference/mean": 0.05213358998298645,
+      "step": 105,
+      "step_time": 10.497982850998596
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.576000239700079,
+      "epoch": 0.00106,
+      "grad_norm": 0.0021153499837964773,
+      "kl": 0.6080902330577374,
+      "learning_rate": 7.999998185825468e-06,
+      "loss": -0.0001,
+      "step": 106,
+      "step_time": 5.98566820799897
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5625,
+      "completions/mean_terminated_length": 2.5625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7922704741358757,
+      "epoch": 0.00107,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.001570796244777739,
+      "kl": 0.8546689227223396,
+      "learning_rate": 7.999998133621676e-06,
+      "loss": -0.0001,
+      "num_tokens": 2569391.0,
+      "reward": 1.7937500476837158,
+      "reward_std": 0.9540871381759644,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5625,
+      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.875,
+      "rewards/probe_shaping_dominance/std": 0.33601075410842896,
+      "rewards/probe_terminal_raw/mean": 0.125,
+      "rewards/probe_terminal_raw/std": 0.33601075410842896,
+      "rewards/rollout_reward_func/mean": -0.75,
+      "rewards/rollout_reward_func/std": 0.6720215082168579,
+      "sampling/importance_sampling_ratio/max": 1.3402096033096313,
+      "sampling/importance_sampling_ratio/mean": 0.863715410232544,
+      "sampling/importance_sampling_ratio/min": 0.29634997248649597,
+      "sampling/sampling_logp_difference/max": 0.9534087181091309,
+      "sampling/sampling_logp_difference/mean": 0.12255297601222992,
+      "step": 107,
+      "step_time": 10.240153937998912
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7928206250071526,
+      "epoch": 0.00108,
+      "grad_norm": 0.0015277264174073935,
+      "kl": 0.7050070539116859,
+      "learning_rate": 7.999998080677404e-06,
+      "loss": -0.0001,
+      "step": 108,
+      "step_time": 6.46414371100218
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.8125,
+      "completions/mean_terminated_length": 2.8125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.5738925524055958,
+      "epoch": 0.00109,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0044546849094331264,
+      "kl": 0.8377486700192094,
+      "learning_rate": 7.999998026992654e-06,
+      "loss": -0.0001,
+      "num_tokens": 2615756.0,
+      "reward": 2.075000047683716,
+      "reward_std": 0.9069623351097107,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.8125,
+      "rewards/probe_completion_length/std": 0.3965577781200409,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.84375,
+      "rewards/probe_shaping_dominance/std": 0.3689020276069641,
+      "rewards/probe_terminal_raw/mean": 0.15625,
+      "rewards/probe_terminal_raw/std": 0.3689020276069641,
+      "rewards/rollout_reward_func/mean": -0.6875,
+      "rewards/rollout_reward_func/std": 0.7378040552139282,
+      "sampling/importance_sampling_ratio/max": 2.154284954071045,
+      "sampling/importance_sampling_ratio/mean": 0.9438944458961487,
+      "sampling/importance_sampling_ratio/min": 0.28246164321899414,
+      "sampling/sampling_logp_difference/max": 0.8599565029144287,
+      "sampling/sampling_logp_difference/mean": 0.126668319106102,
+      "step": 109,
+      "step_time": 10.330337892997704
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.033333334140479565,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.033333334140479565,
+      "entropy": 0.5660623703151941,
+      "epoch": 0.0011,
+      "grad_norm": 0.00280306045897305,
+      "kl": 0.9475272796116769,
+      "learning_rate": 7.999997972567424e-06,
+      "loss": -0.0001,
+      "step": 110,
+      "step_time": 5.766006027999538
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.71875,
+      "completions/mean_terminated_length": 2.71875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6521939821541309,
+      "epoch": 0.00111,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0037616584450006485,
+      "kl": 0.6928037863690406,
+      "learning_rate": 7.999997917401717e-06,
+      "loss": -0.0,
+      "num_tokens": 2660029.0,
+      "reward": 2.543750047683716,
+      "reward_std": 1.3163825273513794,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.71875,
+      "rewards/probe_completion_length/std": 0.45680341124534607,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5625,
+      "rewards/probe_shaping_dominance/std": 0.504016101360321,
+      "rewards/probe_terminal_raw/mean": 0.4375,
+      "rewards/probe_terminal_raw/std": 0.504016101360321,
+      "rewards/rollout_reward_func/mean": -0.125,
+      "rewards/rollout_reward_func/std": 1.008032202720642,
+      "sampling/importance_sampling_ratio/max": 2.5399951934814453,
+      "sampling/importance_sampling_ratio/mean": 0.912597119808197,
+      "sampling/importance_sampling_ratio/min": 0.3494853973388672,
+      "sampling/sampling_logp_difference/max": 1.6019632816314697,
+      "sampling/sampling_logp_difference/mean": 0.11006368696689606,
+      "step": 111,
+      "step_time": 10.330967569998393
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.02083333395421505,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.02083333395421505,
+      "entropy": 0.653257854282856,
+      "epoch": 0.00112,
+      "grad_norm": 0.00019386877829674631,
+      "kl": 0.7184119247831404,
+      "learning_rate": 7.99999786149553e-06,
+      "loss": -0.0,
+      "step": 112,
+      "step_time": 6.43034187000012
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.6875,
+      "completions/mean_terminated_length": 2.6875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.5557065382599831,
+      "epoch": 0.00113,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0019610777962952852,
+      "kl": 0.5786571955832187,
+      "learning_rate": 7.999997804848863e-06,
+      "loss": -0.0001,
+      "num_tokens": 2706298.0,
+      "reward": 2.512500047683716,
+      "reward_std": 1.342542052268982,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.6875,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5625,
+      "rewards/probe_shaping_dominance/std": 0.504016101360321,
+      "rewards/probe_terminal_raw/mean": 0.4375,
+      "rewards/probe_terminal_raw/std": 0.504016101360321,
+      "rewards/rollout_reward_func/mean": -0.125,
+      "rewards/rollout_reward_func/std": 1.008032202720642,
+      "sampling/importance_sampling_ratio/max": 1.146275520324707,
+      "sampling/importance_sampling_ratio/mean": 0.9030254483222961,
+      "sampling/importance_sampling_ratio/min": 0.18932893872261047,
+      "sampling/sampling_logp_difference/max": 1.3923919200897217,
+      "sampling/sampling_logp_difference/mean": 0.06794135272502899,
+      "step": 113,
+      "step_time": 10.613757418001114
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "entropy": 0.5475253388285637,
+      "epoch": 0.00114,
+      "grad_norm": 0.0020037624053657055,
+      "kl": 0.5677314543281682,
+      "learning_rate": 7.999997747461717e-06,
+      "loss": -0.0001,
+      "step": 114,
+      "step_time": 5.993655662001402
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5,
+      "completions/mean_terminated_length": 2.5,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7513568848371506,
+      "epoch": 0.00115,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0003399779088795185,
+      "kl": 0.2799269228707999,
+      "learning_rate": 7.999997689334094e-06,
+      "loss": 0.0,
+      "num_tokens": 2758680.0,
+      "reward": 1.4500000476837158,
+      "reward_std": 0.5080004930496216,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5,
+      "rewards/probe_completion_length/std": 0.5080004930496216,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.589788794517517,
+      "sampling/importance_sampling_ratio/mean": 0.8488104939460754,
+      "sampling/importance_sampling_ratio/min": 0.29667946696281433,
+      "sampling/sampling_logp_difference/max": 1.1271950006484985,
+      "sampling/sampling_logp_difference/mean": 0.11992794275283813,
+      "step": 115,
+      "step_time": 10.530936768000174
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7374518886208534,
+      "epoch": 0.00116,
+      "grad_norm": 0.0005056450027041137,
+      "kl": 0.3286844752728939,
+      "learning_rate": 7.99999763046599e-06,
+      "loss": 0.0,
+      "step": 116,
+      "step_time": 6.617304399002023
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.53125,
+      "completions/mean_terminated_length": 2.53125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8363222517073154,
+      "epoch": 0.00117,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 5.731517740059644e-05,
+      "kl": 0.5686155576258898,
+      "learning_rate": 7.999997570857409e-06,
+      "loss": 0.0,
+      "num_tokens": 2807756.0,
+      "reward": 1.9812500476837158,
+      "reward_std": 1.2309025526046753,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.53125,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.3529267311096191,
+      "sampling/importance_sampling_ratio/mean": 0.8506964445114136,
+      "sampling/importance_sampling_ratio/min": 0.4534793198108673,
+      "sampling/sampling_logp_difference/max": 0.5722330212593079,
+      "sampling/sampling_logp_difference/mean": 0.0960143581032753,
+      "step": 117,
+      "step_time": 10.560414547995606
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8329844176769257,
+      "epoch": 0.00118,
+      "grad_norm": 7.695100794080645e-05,
+      "kl": 0.5957790687680244,
+      "learning_rate": 7.999997510508348e-06,
+      "loss": 0.0,
+      "step": 118,
+      "step_time": 6.05941285299923
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.46875,
+      "completions/mean_terminated_length": 2.46875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7654928229749203,
+      "epoch": 0.00119,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.001598424045369029,
+      "kl": 0.6253797174431384,
+      "learning_rate": 7.999997449418809e-06,
+      "loss": -0.0001,
+      "num_tokens": 2856512.0,
+      "reward": 1.7312500476837158,
+      "reward_std": 1.0696510076522827,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.46875,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.84375,
+      "rewards/probe_shaping_dominance/std": 0.3689020276069641,
+      "rewards/probe_terminal_raw/mean": 0.15625,
+      "rewards/probe_terminal_raw/std": 0.3689020276069641,
+      "rewards/rollout_reward_func/mean": -0.6875,
+      "rewards/rollout_reward_func/std": 0.7378040552139282,
+      "sampling/importance_sampling_ratio/max": 1.546105146408081,
+      "sampling/importance_sampling_ratio/mean": 0.8397697806358337,
+      "sampling/importance_sampling_ratio/min": 0.24415840208530426,
+      "sampling/sampling_logp_difference/max": 1.3526235818862915,
+      "sampling/sampling_logp_difference/mean": 0.12898343801498413,
+      "step": 119,
+      "step_time": 10.996136936004405
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.033333334140479565,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.033333334140479565,
+      "entropy": 0.7672837302088737,
+      "epoch": 0.0012,
+      "grad_norm": 0.0002433511835988611,
+      "kl": 0.6368538588285446,
+      "learning_rate": 7.99999738758879e-06,
+      "loss": -0.0001,
+      "step": 120,
+      "step_time": 6.0576140050015965
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.6875,
+      "completions/mean_terminated_length": 2.6875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.40811267495155334,
+      "epoch": 0.00121,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.002180876210331917,
+      "kl": 1.549589341506362,
+      "learning_rate": 7.999997325018293e-06,
+      "loss": 0.0,
+      "num_tokens": 2900669.0,
+      "reward": 2.668750047683716,
+      "reward_std": 1.3496564626693726,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.6875,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.5,
+      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
+      "rewards/probe_terminal_raw/mean": 0.5,
+      "rewards/probe_terminal_raw/std": 0.5080004930496216,
+      "rewards/rollout_reward_func/mean": 0.0,
+      "rewards/rollout_reward_func/std": 1.0160009860992432,
+      "sampling/importance_sampling_ratio/max": 1.735442876815796,
+      "sampling/importance_sampling_ratio/mean": 0.9358066320419312,
+      "sampling/importance_sampling_ratio/min": 0.05085606127977371,
+      "sampling/sampling_logp_difference/max": 2.797534942626953,
+      "sampling/sampling_logp_difference/mean": 0.151044100522995,
+      "step": 121,
+      "step_time": 10.245774637998693
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.41215409338474274,
+      "epoch": 0.00122,
+      "grad_norm": 0.0017760515911504626,
+      "kl": 1.5261753425002098,
+      "learning_rate": 7.999997261707317e-06,
+      "loss": 0.0,
+      "step": 122,
+      "step_time": 6.374105552997207
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.84375,
+      "completions/mean_terminated_length": 2.84375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.43246884644031525,
+      "epoch": 0.00123,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0008055871003307402,
+      "kl": 1.0272779632359743,
+      "learning_rate": 7.999997197655861e-06,
+      "loss": -0.0001,
+      "num_tokens": 2948581.0,
+      "reward": 2.856250047683716,
+      "reward_std": 1.2276222705841064,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.84375,
+      "rewards/probe_completion_length/std": 0.3689020276069641,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.46875,
+      "rewards/probe_shaping_dominance/std": 0.507007360458374,
+      "rewards/probe_terminal_raw/mean": 0.53125,
+      "rewards/probe_terminal_raw/std": 0.507007360458374,
+      "rewards/rollout_reward_func/mean": 0.0625,
+      "rewards/rollout_reward_func/std": 1.014014720916748,
+      "sampling/importance_sampling_ratio/max": 1.839868426322937,
+      "sampling/importance_sampling_ratio/mean": 0.8939834833145142,
+      "sampling/importance_sampling_ratio/min": 0.07948717474937439,
+      "sampling/sampling_logp_difference/max": 2.2706401348114014,
+      "sampling/sampling_logp_difference/mean": 0.11110737919807434,
+      "step": 123,
+      "step_time": 10.275854114002868
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.43315667286515236,
+      "epoch": 0.00124,
+      "grad_norm": 0.0008163482998497784,
+      "kl": 1.0497613418847322,
+      "learning_rate": 7.999997132863928e-06,
+      "loss": -0.0001,
+      "step": 124,
+      "step_time": 5.97916301999976
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.3833901360630989,
+      "epoch": 0.00125,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0007064693490974605,
+      "kl": 1.3148364201188087,
+      "learning_rate": 7.999997067331516e-06,
+      "loss": 0.0,
+      "num_tokens": 2996178.0,
+      "reward": 2.231250047683716,
+      "reward_std": 1.0846248865127563,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.78125,
+      "rewards/probe_completion_length/std": 0.420013427734375,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.0497198104858398,
+      "sampling/importance_sampling_ratio/mean": 0.7578398585319519,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 2.767157554626465,
+      "sampling/sampling_logp_difference/mean": 0.1752256155014038,
+      "step": 125,
+      "step_time": 10.13536022900189
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.38512268476188183,
+      "epoch": 0.00126,
+      "grad_norm": 0.000761943229008466,
+      "kl": 1.3329130411148071,
+      "learning_rate": 7.999997001058625e-06,
+      "loss": 0.0,
+      "step": 126,
+      "step_time": 6.731002937000085
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.40625,
+      "completions/mean_terminated_length": 2.40625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.669311985373497,
+      "epoch": 0.00127,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00017574324738234282,
+      "kl": 0.6046575158834457,
+      "learning_rate": 7.999996934045254e-06,
+      "loss": 0.0,
+      "num_tokens": 3041147.0,
+      "reward": 1.8562500476837158,
+      "reward_std": 1.2790968418121338,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.40625,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.3875808715820312,
+      "sampling/importance_sampling_ratio/mean": 0.9241994619369507,
+      "sampling/importance_sampling_ratio/min": 0.7366468906402588,
+      "sampling/sampling_logp_difference/max": 0.3816380500793457,
+      "sampling/sampling_logp_difference/mean": 0.05508778989315033,
+      "step": 127,
+      "step_time": 10.081132162002177
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6669952869415283,
+      "epoch": 0.00128,
+      "grad_norm": 0.0002272838755743578,
+      "kl": 0.62955535762012,
+      "learning_rate": 7.999996866291406e-06,
+      "loss": 0.0,
+      "step": 128,
+      "step_time": 5.974870206999185
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.59375,
+      "completions/mean_terminated_length": 2.59375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.564176794141531,
+      "epoch": 0.00129,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0013212214689701796,
+      "kl": 0.9751707948744297,
+      "learning_rate": 7.999996797797079e-06,
+      "loss": -0.0,
+      "num_tokens": 3086335.0,
+      "reward": 2.231250047683716,
+      "reward_std": 1.3255400657653809,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.59375,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.65625,
+      "rewards/probe_shaping_dominance/std": 0.4825586974620819,
+      "rewards/probe_terminal_raw/mean": 0.34375,
+      "rewards/probe_terminal_raw/std": 0.4825586974620819,
+      "rewards/rollout_reward_func/mean": -0.3125,
+      "rewards/rollout_reward_func/std": 0.9651173949241638,
+      "sampling/importance_sampling_ratio/max": 1.2364485263824463,
+      "sampling/importance_sampling_ratio/mean": 0.8747109174728394,
+      "sampling/importance_sampling_ratio/min": 0.2390734702348709,
+      "sampling/sampling_logp_difference/max": 1.3546556234359741,
+      "sampling/sampling_logp_difference/mean": 0.07768772542476654,
+      "step": 129,
+      "step_time": 10.650671339002656
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.5586648434400558,
+      "epoch": 0.0013,
+      "grad_norm": 0.001662288443185389,
+      "kl": 0.9744510091841221,
+      "learning_rate": 7.999996728562273e-06,
+      "loss": -0.0,
+      "step": 130,
+      "step_time": 6.473435007999797
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.012500000186264515,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.75,
+      "completions/mean_terminated_length": 2.75,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.3778875097632408,
+      "epoch": 0.00131,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0024475299287587404,
+      "kl": 1.603209212422371,
+      "learning_rate": 7.999996658586989e-06,
+      "loss": -0.0001,
+      "num_tokens": 3133159.0,
+      "reward": 2.637500047683716,
+      "reward_std": 1.3060035705566406,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.75,
+      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.53125,
+      "rewards/probe_shaping_dominance/std": 0.507007360458374,
+      "rewards/probe_terminal_raw/mean": 0.46875,
+      "rewards/probe_terminal_raw/std": 0.507007360458374,
+      "rewards/rollout_reward_func/mean": -0.0625,
+      "rewards/rollout_reward_func/std": 1.014014720916748,
+      "sampling/importance_sampling_ratio/max": 1.2442551851272583,
+      "sampling/importance_sampling_ratio/mean": 0.8854572772979736,
+      "sampling/importance_sampling_ratio/min": 0.10244226455688477,
+      "sampling/sampling_logp_difference/max": 2.1706011295318604,
+      "sampling/sampling_logp_difference/mean": 0.1017080694437027,
+      "step": 131,
+      "step_time": 10.024731007997616
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.012500000186264515,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.37728825211524963,
+      "epoch": 0.00132,
+      "grad_norm": 0.0006488582585006952,
+      "kl": 1.5983238890767097,
+      "learning_rate": 7.999996587871225e-06,
+      "loss": -0.0001,
+      "step": 132,
+      "step_time": 6.3365548879992275
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6743998825550079,
+      "epoch": 0.00133,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.006993595510721207,
+      "kl": 3.7676026318222284,
+      "learning_rate": 7.999996516414982e-06,
+      "loss": -0.0002,
+      "num_tokens": 3186308.0,
+      "reward": 1.8250000476837158,
+      "reward_std": 0.941858172416687,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.875,
+      "rewards/probe_shaping_dominance/std": 0.33601075410842896,
+      "rewards/probe_terminal_raw/mean": 0.125,
+      "rewards/probe_terminal_raw/std": 0.33601075410842896,
+      "rewards/rollout_reward_func/mean": -0.75,
+      "rewards/rollout_reward_func/std": 0.6720215082168579,
+      "sampling/importance_sampling_ratio/max": 2.262389898300171,
+      "sampling/importance_sampling_ratio/mean": 0.9818167686462402,
+      "sampling/importance_sampling_ratio/min": 0.02067597396671772,
+      "sampling/sampling_logp_difference/max": 3.6869354248046875,
+      "sampling/sampling_logp_difference/mean": 0.22966735064983368,
+      "step": 133,
+      "step_time": 11.05697975199837
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6868020631372929,
+      "epoch": 0.00134,
+      "grad_norm": 0.0053808679804205894,
+      "kl": 2.8858383037149906,
+      "learning_rate": 7.999996444218262e-06,
+      "loss": -0.0002,
+      "step": 134,
+      "step_time": 6.313005753003381
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 6.0,
+      "completions/max_terminated_length": 6.0,
+      "completions/mean_length": 2.40625,
+      "completions/mean_terminated_length": 2.40625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6054645292460918,
+      "epoch": 0.00135,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 3.4668184525799006e-05,
+      "kl": 0.07004273735219613,
+      "learning_rate": 7.999996371281063e-06,
+      "loss": 0.0,
+      "num_tokens": 3232062.0,
+      "reward": 1.8249999284744263,
+      "reward_std": 1.4535826444625854,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.375,
+      "rewards/probe_completion_length/std": 0.6599119901657104,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.3307595252990723,
+      "sampling/importance_sampling_ratio/mean": 0.9332914352416992,
+      "sampling/importance_sampling_ratio/min": 0.5020673274993896,
+      "sampling/sampling_logp_difference/max": 0.3572865128517151,
+      "sampling/sampling_logp_difference/mean": 0.05530416965484619,
+      "step": 135,
+      "step_time": 10.475520918997063
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.610666386783123,
+      "epoch": 0.00136,
+      "grad_norm": 3.350508268340491e-05,
+      "kl": 0.06979758548550308,
+      "learning_rate": 7.999996297603385e-06,
+      "loss": 0.0,
+      "step": 136,
+      "step_time": 6.031642523999835
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.46875,
+      "completions/mean_terminated_length": 2.46875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6555749177932739,
+      "epoch": 0.00137,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0013054372975602746,
+      "kl": 0.5964277469320223,
+      "learning_rate": 7.999996223185228e-06,
+      "loss": -0.0001,
+      "num_tokens": 3284053.0,
+      "reward": 1.7937500476837158,
+      "reward_std": 1.1390254497528076,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.46875,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.8125,
+      "rewards/probe_shaping_dominance/std": 0.3965577781200409,
+      "rewards/probe_terminal_raw/mean": 0.1875,
+      "rewards/probe_terminal_raw/std": 0.3965577781200409,
+      "rewards/rollout_reward_func/mean": -0.625,
+      "rewards/rollout_reward_func/std": 0.7931155562400818,
+      "sampling/importance_sampling_ratio/max": 1.1609679460525513,
+      "sampling/importance_sampling_ratio/mean": 0.8882247805595398,
+      "sampling/importance_sampling_ratio/min": 0.36374035477638245,
+      "sampling/sampling_logp_difference/max": 0.9692283272743225,
+      "sampling/sampling_logp_difference/mean": 0.064999520778656,
+      "step": 137,
+      "step_time": 10.692799777998516
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6570683643221855,
+      "epoch": 0.00138,
+      "grad_norm": 0.0015739274676889181,
+      "kl": 0.5917147692525759,
+      "learning_rate": 7.999996148026594e-06,
+      "loss": -0.0001,
+      "step": 138,
+      "step_time": 6.068219756996768
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6512251161038876,
+      "epoch": 0.00139,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00028360536089167,
+      "kl": 0.5178780551068485,
+      "learning_rate": 7.99999607212748e-06,
+      "loss": 0.0,
+      "num_tokens": 3334405.0,
+      "reward": 1.6062500476837158,
+      "reward_std": 0.4825586974620819,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.65625,
+      "rewards/probe_completion_length/std": 0.4825586974620819,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.6509288549423218,
+      "sampling/importance_sampling_ratio/mean": 0.893890380859375,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 2.125434160232544,
+      "sampling/sampling_logp_difference/mean": 0.13162320852279663,
+      "step": 139,
+      "step_time": 10.860906833999252
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6622384190559387,
+      "epoch": 0.0014,
+      "grad_norm": 0.0002564970636740327,
+      "kl": 0.4972880659624934,
+      "learning_rate": 7.999995995487888e-06,
+      "loss": 0.0,
+      "step": 140,
+      "step_time": 6.4930414069986
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6598617434501648,
+      "epoch": 0.00141,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.001187963760457933,
+      "kl": 0.5750172566622496,
+      "learning_rate": 7.999995918107818e-06,
+      "loss": -0.0001,
+      "num_tokens": 3383638.0,
+      "reward": 2.043750047683716,
+      "reward_std": 1.1175830364227295,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.65625,
+      "rewards/probe_completion_length/std": 0.4825586974620819,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.3601293563842773,
+      "sampling/importance_sampling_ratio/mean": 0.8598825931549072,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 1.3208266496658325,
+      "sampling/sampling_logp_difference/mean": 0.11825218796730042,
+      "step": 141,
+      "step_time": 10.291784546998315
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6577947102487087,
+      "epoch": 0.00142,
+      "grad_norm": 0.0006299904780462384,
+      "kl": 0.5920837037265301,
+      "learning_rate": 7.999995839987269e-06,
+      "loss": -0.0001,
+      "step": 142,
+      "step_time": 6.125413964000472
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 7.0,
+      "completions/max_terminated_length": 7.0,
+      "completions/mean_length": 2.5,
+      "completions/mean_terminated_length": 2.5,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8532420322299004,
+      "epoch": 0.00143,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.233860767679289e-05,
+      "kl": 0.057465479942038655,
+      "learning_rate": 7.999995761126243e-06,
+      "loss": 0.0,
+      "num_tokens": 3432038.0,
+      "reward": 1.5125000476837158,
+      "reward_std": 1.2684128284454346,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5625,
+      "rewards/probe_completion_length/std": 1.268412709236145,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.3138529062271118,
+      "sampling/importance_sampling_ratio/mean": 0.8958837985992432,
+      "sampling/importance_sampling_ratio/min": 0.5682582259178162,
+      "sampling/sampling_logp_difference/max": 0.375757098197937,
+      "sampling/sampling_logp_difference/mean": 0.06977839767932892,
+      "step": 143,
+      "step_time": 10.140005126995675
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8545014560222626,
+      "epoch": 0.00144,
+      "grad_norm": 7.701283902861178e-05,
+      "kl": 0.06713397847488523,
+      "learning_rate": 7.999995681524736e-06,
+      "loss": 0.0,
+      "step": 144,
+      "step_time": 6.440495836997798
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.4375,
+      "completions/mean_terminated_length": 2.4375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8347389176487923,
+      "epoch": 0.00145,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 7.582701073260978e-05,
+      "kl": 0.21212944004219025,
+      "learning_rate": 7.999995601182752e-06,
+      "loss": 0.0,
+      "num_tokens": 3482547.0,
+      "reward": 1.3875000476837158,
+      "reward_std": 0.504016101360321,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.4375,
+      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.8214629888534546,
+      "sampling/importance_sampling_ratio/mean": 0.9022046327590942,
+      "sampling/importance_sampling_ratio/min": 0.45289722084999084,
+      "sampling/sampling_logp_difference/max": 0.6827032566070557,
+      "sampling/sampling_logp_difference/mean": 0.10910705476999283,
+      "step": 145,
+      "step_time": 10.462476491999041
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8556515723466873,
+      "epoch": 0.00146,
+      "grad_norm": 7.53724089008756e-05,
+      "kl": 0.21599587588571012,
+      "learning_rate": 7.999995520100289e-06,
+      "loss": 0.0,
+      "step": 146,
+      "step_time": 6.58235243099989
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.40625,
+      "completions/mean_terminated_length": 2.40625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8366929292678833,
+      "epoch": 0.00147,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0003870174987241626,
+      "kl": 0.24507740326225758,
+      "learning_rate": 7.999995438277348e-06,
+      "loss": 0.0,
+      "num_tokens": 3528662.0,
+      "reward": 1.3875000476837158,
+      "reward_std": 0.504016101360321,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.40625,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.3872897624969482,
+      "sampling/importance_sampling_ratio/mean": 0.8565528392791748,
+      "sampling/importance_sampling_ratio/min": 0.1745889037847519,
+      "sampling/sampling_logp_difference/max": 1.6216473579406738,
+      "sampling/sampling_logp_difference/mean": 0.13993701338768005,
+      "step": 147,
+      "step_time": 10.136645781998595
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8375646397471428,
+      "epoch": 0.00148,
+      "grad_norm": 0.0006000733701512218,
+      "kl": 0.3029486406594515,
+      "learning_rate": 7.99999535571393e-06,
+      "loss": 0.0,
+      "step": 148,
+      "step_time": 6.332877236998684
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5625,
+      "completions/mean_terminated_length": 2.5625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7925528138875961,
+      "epoch": 0.00149,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 8.313127182191238e-05,
+      "kl": 0.7751652393490076,
+      "learning_rate": 7.999995272410032e-06,
+      "loss": 0.0,
+      "num_tokens": 3578208.0,
+      "reward": 2.012500047683716,
+      "reward_std": 1.2164862155914307,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5625,
+      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.291900873184204,
+      "sampling/importance_sampling_ratio/mean": 0.8923901319503784,
+      "sampling/importance_sampling_ratio/min": 0.5138108134269714,
+      "sampling/sampling_logp_difference/max": 0.47057437896728516,
+      "sampling/sampling_logp_difference/mean": 0.08149246871471405,
+      "step": 149,
+      "step_time": 10.504922516000079
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8027943335473537,
+      "epoch": 0.0015,
+      "grad_norm": 5.7398243370698765e-05,
+      "kl": 0.7740726545453072,
+      "learning_rate": 7.999995188365656e-06,
+      "loss": 0.0,
+      "step": 150,
+      "step_time": 6.047526541997286
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.5021162088960409,
+      "epoch": 0.00151,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 4.203694697935134e-05,
+      "kl": 0.8826476074755192,
+      "learning_rate": 7.999995103580802e-06,
+      "loss": 0.0,
+      "num_tokens": 3622385.0,
+      "reward": 2.575000047683716,
+      "reward_std": 1.4312187433242798,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5,
+      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
+      "rewards/probe_terminal_raw/mean": 0.5,
+      "rewards/probe_terminal_raw/std": 0.5080004930496216,
+      "rewards/rollout_reward_func/mean": 0.0,
+      "rewards/rollout_reward_func/std": 1.0160009860992432,
+      "sampling/importance_sampling_ratio/max": 1.629455804824829,
+      "sampling/importance_sampling_ratio/mean": 0.9585285186767578,
+      "sampling/importance_sampling_ratio/min": 0.703690767288208,
+      "sampling/sampling_logp_difference/max": 0.5472321510314941,
+      "sampling/sampling_logp_difference/mean": 0.040512826293706894,
+      "step": 151,
+      "step_time": 10.376955908001037
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.501195352524519,
+      "epoch": 0.00152,
+      "grad_norm": 4.0065049688564613e-05,
+      "kl": 0.8829234223812819,
+      "learning_rate": 7.99999501805547e-06,
+      "loss": 0.0,
+      "step": 152,
+      "step_time": 5.959771870997429
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6162157356739044,
+      "epoch": 0.00153,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 8.145786705426872e-05,
+      "kl": 0.8375249132514,
+      "learning_rate": 7.99999493178966e-06,
+      "loss": 0.0,
+      "num_tokens": 3668995.0,
+      "reward": 2.231250047683716,
+      "reward_std": 1.0846248865127563,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.78125,
+      "rewards/probe_completion_length/std": 0.420013427734375,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 2.3910458087921143,
+      "sampling/importance_sampling_ratio/mean": 0.8976485729217529,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 1.0475459098815918,
+      "sampling/sampling_logp_difference/mean": 0.1141430139541626,
+      "step": 153,
+      "step_time": 10.604804011001761
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6082192212343216,
+      "epoch": 0.00154,
+      "grad_norm": 9.017075353767723e-05,
+      "kl": 0.8417555466294289,
+      "learning_rate": 7.99999484478337e-06,
+      "loss": 0.0,
+      "step": 154,
+      "step_time": 6.000888992999535
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.875,
+      "completions/mean_terminated_length": 2.875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.4833748936653137,
+      "epoch": 0.00155,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0174336489289999,
+      "kl": 12.712480776011944,
+      "learning_rate": 7.999994757036603e-06,
+      "loss": 0.0001,
+      "num_tokens": 3714734.0,
+      "reward": 2.637500047683716,
+      "reward_std": 1.1482806205749512,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.875,
+      "rewards/probe_completion_length/std": 0.33601075410842896,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.59375,
+      "rewards/probe_shaping_dominance/std": 0.49899089336395264,
+      "rewards/probe_terminal_raw/mean": 0.40625,
+      "rewards/probe_terminal_raw/std": 0.49899089336395264,
+      "rewards/rollout_reward_func/mean": -0.1875,
+      "rewards/rollout_reward_func/std": 0.9979817867279053,
+      "sampling/importance_sampling_ratio/max": 2.032945394515991,
+      "sampling/importance_sampling_ratio/mean": 0.9355040192604065,
+      "sampling/importance_sampling_ratio/min": 0.22561033070087433,
+      "sampling/sampling_logp_difference/max": 1.1889557838439941,
+      "sampling/sampling_logp_difference/mean": 0.11608728021383286,
+      "step": 155,
+      "step_time": 10.715239364000809
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.4995018355548382,
+      "epoch": 0.00156,
+      "grad_norm": 0.003404801245778799,
+      "kl": 5.325677994638681,
+      "learning_rate": 7.999994668549356e-06,
+      "loss": 0.0,
+      "step": 156,
+      "step_time": 6.548639589003869
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5,
+      "completions/mean_terminated_length": 2.5,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9601438865065575,
+      "epoch": 0.00157,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0019437572918832302,
+      "kl": 0.6555760141345672,
+      "learning_rate": 7.999994579321633e-06,
+      "loss": -0.0001,
+      "num_tokens": 3760285.0,
+      "reward": 1.7000000476837158,
+      "reward_std": 0.9837387204170227,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5,
+      "rewards/probe_completion_length/std": 0.5080004930496216,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.875,
+      "rewards/probe_shaping_dominance/std": 0.33601075410842896,
+      "rewards/probe_terminal_raw/mean": 0.125,
+      "rewards/probe_terminal_raw/std": 0.33601075410842896,
+      "rewards/rollout_reward_func/mean": -0.75,
+      "rewards/rollout_reward_func/std": 0.6720215082168579,
+      "sampling/importance_sampling_ratio/max": 2.429400682449341,
+      "sampling/importance_sampling_ratio/mean": 0.8863468170166016,
+      "sampling/importance_sampling_ratio/min": 0.2631182074546814,
+      "sampling/sampling_logp_difference/max": 0.9481115341186523,
+      "sampling/sampling_logp_difference/mean": 0.1506330370903015,
+      "step": 157,
+      "step_time": 10.192884709997088
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9771413132548332,
+      "epoch": 0.00158,
+      "grad_norm": 0.0026333045680075884,
+      "kl": 0.6077049063460436,
+      "learning_rate": 7.999994489353432e-06,
+      "loss": -0.0001,
+      "step": 158,
+      "step_time": 6.394098832000964
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 7.0,
+      "completions/max_terminated_length": 7.0,
+      "completions/mean_length": 2.96875,
+      "completions/mean_terminated_length": 2.96875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.5719749964773655,
+      "epoch": 0.00159,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 8.808910934021696e-05,
+      "kl": 0.38396941107930616,
+      "learning_rate": 7.999994398644752e-06,
+      "loss": 0.0,
+      "num_tokens": 3810633.0,
+      "reward": 2.4812498092651367,
+      "reward_std": 1.43649160861969,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 2.03125,
+      "rewards/probe_completion_length/std": 1.1495966911315918,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.761324167251587,
+      "sampling/importance_sampling_ratio/mean": 0.8197870850563049,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 1.5803544521331787,
+      "sampling/sampling_logp_difference/mean": 0.0992826372385025,
+      "step": 159,
+      "step_time": 10.46122989000105
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.574010081589222,
+      "epoch": 0.0016,
+      "grad_norm": 7.729078060947359e-05,
+      "kl": 0.3841613693512045,
+      "learning_rate": 7.999994307195594e-06,
+      "loss": 0.0,
+      "step": 160,
+      "step_time": 6.568957989997216
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.6875,
+      "completions/mean_terminated_length": 2.6875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6980656832456589,
+      "epoch": 0.00161,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0001728427450871095,
+      "kl": 0.9108022870495915,
+      "learning_rate": 7.99999421500596e-06,
+      "loss": 0.0,
+      "num_tokens": 3862246.0,
+      "reward": 2.168750047683716,
+      "reward_std": 1.1283552646636963,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.6875,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.7241743803024292,
+      "sampling/importance_sampling_ratio/mean": 0.8921574354171753,
+      "sampling/importance_sampling_ratio/min": 0.5468968152999878,
+      "sampling/sampling_logp_difference/max": 0.5729777812957764,
+      "sampling/sampling_logp_difference/mean": 0.09277266263961792,
+      "step": 161,
+      "step_time": 10.87614541900075
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7105224430561066,
+      "epoch": 0.00162,
+      "grad_norm": 0.00016276795940939337,
+      "kl": 0.9079902372322977,
+      "learning_rate": 7.999994122075845e-06,
+      "loss": 0.0,
+      "step": 162,
+      "step_time": 6.721154226997896
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 8.0,
+      "completions/max_terminated_length": 8.0,
+      "completions/mean_length": 2.53125,
+      "completions/mean_terminated_length": 2.53125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9394364431500435,
+      "epoch": 0.00163,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.971744733164087e-05,
+      "kl": 0.12017604021821171,
+      "learning_rate": 7.999994028405253e-06,
+      "loss": 0.0,
+      "num_tokens": 3908510.0,
+      "reward": 1.5437500476837158,
+      "reward_std": 1.433681845664978,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.59375,
+      "rewards/probe_completion_length/std": 1.433681845664978,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.252063512802124,
+      "sampling/importance_sampling_ratio/mean": 0.8833824396133423,
+      "sampling/importance_sampling_ratio/min": 0.5937425494194031,
+      "sampling/sampling_logp_difference/max": 0.39126133918762207,
+      "sampling/sampling_logp_difference/mean": 0.09234651923179626,
+      "step": 163,
+      "step_time": 10.124098319998666
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9440240561962128,
+      "epoch": 0.00164,
+      "grad_norm": 5.371127190301195e-05,
+      "kl": 0.11512814310844988,
+      "learning_rate": 7.999993933994183e-06,
+      "loss": 0.0,
+      "step": 164,
+      "step_time": 6.441034747997037
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.90625,
+      "completions/mean_terminated_length": 2.90625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7190432026982307,
+      "epoch": 0.00165,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.004245645832270384,
+      "kl": 0.8631834057159722,
+      "learning_rate": 7.999993838842636e-06,
+      "loss": 0.0,
+      "num_tokens": 3957494.0,
+      "reward": 2.231250047683716,
+      "reward_std": 0.8884337544441223,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.90625,
+      "rewards/probe_completion_length/std": 0.2961445748806,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.8125,
+      "rewards/probe_shaping_dominance/std": 0.3965577781200409,
+      "rewards/probe_terminal_raw/mean": 0.1875,
+      "rewards/probe_terminal_raw/std": 0.3965577781200409,
+      "rewards/rollout_reward_func/mean": -0.625,
+      "rewards/rollout_reward_func/std": 0.7931155562400818,
+      "sampling/importance_sampling_ratio/max": 1.784828782081604,
+      "sampling/importance_sampling_ratio/mean": 0.9233963489532471,
+      "sampling/importance_sampling_ratio/min": 0.45206698775291443,
+      "sampling/sampling_logp_difference/max": 0.6415358781814575,
+      "sampling/sampling_logp_difference/mean": 0.0921846255660057,
+      "step": 165,
+      "step_time": 10.395008625999253
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7224007025361061,
+      "epoch": 0.00166,
+      "grad_norm": 0.004814396146684885,
+      "kl": 0.8468241069931537,
+      "learning_rate": 7.99999374295061e-06,
+      "loss": 0.0,
+      "step": 166,
+      "step_time": 6.5217645579959935
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 8.0,
+      "completions/max_terminated_length": 8.0,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.703140377998352,
+      "epoch": 0.00167,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0013086164835840464,
+      "kl": 0.5565936341881752,
+      "learning_rate": 7.999993646318106e-06,
+      "loss": -0.0001,
+      "num_tokens": 4001670.0,
+      "reward": 2.7312498092651367,
+      "reward_std": 1.8091946840286255,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.84375,
+      "rewards/probe_completion_length/std": 1.3937504291534424,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.53125,
+      "rewards/probe_shaping_dominance/std": 0.507007360458374,
+      "rewards/probe_terminal_raw/mean": 0.46875,
+      "rewards/probe_terminal_raw/std": 0.507007360458374,
+      "rewards/rollout_reward_func/mean": -0.0625,
+      "rewards/rollout_reward_func/std": 1.014014720916748,
+      "sampling/importance_sampling_ratio/max": 1.0223342180252075,
+      "sampling/importance_sampling_ratio/mean": 0.8933640718460083,
+      "sampling/importance_sampling_ratio/min": 0.4319377839565277,
+      "sampling/sampling_logp_difference/max": 0.7988812327384949,
+      "sampling/sampling_logp_difference/mean": 0.056503020226955414,
+      "step": 167,
+      "step_time": 10.640347878001194
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7059839554131031,
+      "epoch": 0.00168,
+      "grad_norm": 0.0014117551036179066,
+      "kl": 0.5676930854097009,
+      "learning_rate": 7.999993548945123e-06,
+      "loss": -0.0001,
+      "step": 168,
+      "step_time": 6.005034185000113
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5625,
+      "completions/mean_terminated_length": 2.5625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6460188962519169,
+      "epoch": 0.00169,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.002058417769148946,
+      "kl": 0.8250314202159643,
+      "learning_rate": 7.999993450831664e-06,
+      "loss": -0.0,
+      "num_tokens": 4047467.0,
+      "reward": 2.450000047683716,
+      "reward_std": 1.4591203927993774,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5625,
+      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.53125,
+      "rewards/probe_shaping_dominance/std": 0.507007360458374,
+      "rewards/probe_terminal_raw/mean": 0.46875,
+      "rewards/probe_terminal_raw/std": 0.507007360458374,
+      "rewards/rollout_reward_func/mean": -0.0625,
+      "rewards/rollout_reward_func/std": 1.014014720916748,
+      "sampling/importance_sampling_ratio/max": 1.4465869665145874,
+      "sampling/importance_sampling_ratio/mean": 0.930103600025177,
+      "sampling/importance_sampling_ratio/min": 0.6360700130462646,
+      "sampling/sampling_logp_difference/max": 0.5103611946105957,
+      "sampling/sampling_logp_difference/mean": 0.05351896956562996,
+      "step": 169,
+      "step_time": 10.101505926002574
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.012500000186264515,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.6425128392875195,
+      "epoch": 0.0017,
+      "grad_norm": 0.0008039339445531368,
+      "kl": 0.8195577897131443,
+      "learning_rate": 7.999993351977727e-06,
+      "loss": -0.0,
+      "step": 170,
+      "step_time": 6.393504755000322
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.96875,
+      "completions/mean_terminated_length": 2.96875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.4804557263851166,
+      "epoch": 0.00171,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0005060956464149058,
+      "kl": 1.3053148314356804,
+      "learning_rate": 7.999993252383311e-06,
+      "loss": 0.0,
+      "num_tokens": 4092174.0,
+      "reward": 2.450000047683716,
+      "reward_std": 0.8798826932907104,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.96875,
+      "rewards/probe_completion_length/std": 0.1767766922712326,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.7767850160598755,
+      "sampling/importance_sampling_ratio/mean": 0.9398790597915649,
+      "sampling/importance_sampling_ratio/min": 0.4483863115310669,
+      "sampling/sampling_logp_difference/max": 0.7272882461547852,
+      "sampling/sampling_logp_difference/mean": 0.0852513238787651,
+      "step": 171,
+      "step_time": 10.745308484996713
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.47462591901421547,
+      "epoch": 0.00172,
+      "grad_norm": 0.0004229886981192976,
+      "kl": 1.3004812747240067,
+      "learning_rate": 7.999993152048418e-06,
+      "loss": 0.0,
+      "step": 172,
+      "step_time": 5.9784996139987925
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.6875,
+      "completions/mean_terminated_length": 2.6875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.807799257338047,
+      "epoch": 0.00173,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.002302733715623617,
+      "kl": 1.0078366212546825,
+      "learning_rate": 7.999993050973047e-06,
+      "loss": -0.0,
+      "num_tokens": 4144646.0,
+      "reward": 2.106250047683716,
+      "reward_std": 1.0809009075164795,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.6875,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.0853863954544067,
+      "sampling/importance_sampling_ratio/mean": 0.8166617155075073,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 2.208253860473633,
+      "sampling/sampling_logp_difference/mean": 0.13318130373954773,
+      "step": 173,
+      "step_time": 10.684413701004814
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "entropy": 0.8114959076046944,
+      "epoch": 0.00174,
+      "grad_norm": 0.0003066986973863095,
+      "kl": 1.0182720869779587,
+      "learning_rate": 7.999992949157197e-06,
+      "loss": -0.0,
+      "step": 174,
+      "step_time": 6.4915752000015345
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.75,
+      "completions/mean_terminated_length": 2.75,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.629776481539011,
+      "epoch": 0.00175,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 7.2035203629639e-05,
+      "kl": 1.204437032341957,
+      "learning_rate": 7.999992846600872e-06,
+      "loss": 0.0,
+      "num_tokens": 4193766.0,
+      "reward": 2.700000047683716,
+      "reward_std": 1.319823980331421,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.75,
+      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5,
+      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
+      "rewards/probe_terminal_raw/mean": 0.5,
+      "rewards/probe_terminal_raw/std": 0.5080004930496216,
+      "rewards/rollout_reward_func/mean": 0.0,
+      "rewards/rollout_reward_func/std": 1.0160009860992432,
+      "sampling/importance_sampling_ratio/max": 1.4522444009780884,
+      "sampling/importance_sampling_ratio/mean": 0.8934487700462341,
+      "sampling/importance_sampling_ratio/min": 0.519263505935669,
+      "sampling/sampling_logp_difference/max": 0.5981459617614746,
+      "sampling/sampling_logp_difference/mean": 0.06508005410432816,
+      "step": 175,
+      "step_time": 10.309263899000143
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6329386681318283,
+      "epoch": 0.00176,
+      "grad_norm": 6.6491833422333e-05,
+      "kl": 1.209246039390564,
+      "learning_rate": 7.999992743304069e-06,
+      "loss": 0.0,
+      "step": 176,
+      "step_time": 6.481499650000842
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.75,
+      "completions/mean_terminated_length": 2.75,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8266304433345795,
+      "epoch": 0.00177,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0036097096744924784,
+      "kl": 1.0050737857818604,
+      "learning_rate": 7.999992639266786e-06,
+      "loss": -0.0001,
+      "num_tokens": 4238693.0,
+      "reward": 2.450000047683716,
+      "reward_std": 1.2443419694900513,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.75,
+      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.625,
+      "rewards/probe_shaping_dominance/std": 0.49186936020851135,
+      "rewards/probe_terminal_raw/mean": 0.375,
+      "rewards/probe_terminal_raw/std": 0.49186936020851135,
+      "rewards/rollout_reward_func/mean": -0.25,
+      "rewards/rollout_reward_func/std": 0.9837387204170227,
+      "sampling/importance_sampling_ratio/max": 1.5838106870651245,
+      "sampling/importance_sampling_ratio/mean": 0.8720427751541138,
+      "sampling/importance_sampling_ratio/min": 0.2907257378101349,
+      "sampling/sampling_logp_difference/max": 1.1655175685882568,
+      "sampling/sampling_logp_difference/mean": 0.12054307758808136,
+      "step": 177,
+      "step_time": 10.574092776001635
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "entropy": 0.8272599652409554,
+      "epoch": 0.00178,
+      "grad_norm": 0.0030079202260822058,
+      "kl": 1.0556645542383194,
+      "learning_rate": 7.999992534489026e-06,
+      "loss": -0.0002,
+      "step": 178,
+      "step_time": 5.956353588000638
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.46875,
+      "completions/mean_terminated_length": 2.46875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9563344791531563,
+      "epoch": 0.00179,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.19224582612514496,
+      "kl": 27.414744553156197,
+      "learning_rate": 7.99999242897079e-06,
+      "loss": 0.0003,
+      "num_tokens": 4289683.0,
+      "reward": 1.4812500476837158,
+      "reward_std": 0.507007360458374,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.46875,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 2.3811004161834717,
+      "sampling/importance_sampling_ratio/mean": 0.8373110294342041,
+      "sampling/importance_sampling_ratio/min": 0.09704399853944778,
+      "sampling/sampling_logp_difference/max": 2.7439370155334473,
+      "sampling/sampling_logp_difference/mean": 0.1736406534910202,
+      "step": 179,
+      "step_time": 10.284924712994325
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.006806768476963,
+      "epoch": 0.0018,
+      "grad_norm": 0.00011419090878916904,
+      "kl": 0.5376546527259052,
+      "learning_rate": 7.999992322712075e-06,
+      "loss": 0.0,
+      "step": 180,
+      "step_time": 6.466327140999056
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.53125,
+      "completions/mean_terminated_length": 2.53125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9934074729681015,
+      "epoch": 0.00181,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00010348748764954507,
+      "kl": 0.4358516810461879,
+      "learning_rate": 7.999992215712882e-06,
+      "loss": 0.0,
+      "num_tokens": 4340217.0,
+      "reward": 1.4812500476837158,
+      "reward_std": 0.507007360458374,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.53125,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.6744569540023804,
+      "sampling/importance_sampling_ratio/mean": 0.8735929131507874,
+      "sampling/importance_sampling_ratio/min": 0.33649957180023193,
+      "sampling/sampling_logp_difference/max": 0.7953200340270996,
+      "sampling/sampling_logp_difference/mean": 0.12820304930210114,
+      "step": 181,
+      "step_time": 10.971331862001534
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9945544302463531,
+      "epoch": 0.00182,
+      "grad_norm": 8.246089419117197e-05,
+      "kl": 0.4200055494438857,
+      "learning_rate": 7.999992107973214e-06,
+      "loss": 0.0,
+      "step": 182,
+      "step_time": 6.219140510998841
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7094335816800594,
+      "epoch": 0.00183,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 5.242080078460276e-05,
+      "kl": 0.7885913448408246,
+      "learning_rate": 7.999991999493065e-06,
+      "loss": 0.0,
+      "num_tokens": 4387730.0,
+      "reward": 2.106250047683716,
+      "reward_std": 1.1670026779174805,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.6966081857681274,
+      "sampling/importance_sampling_ratio/mean": 0.8870027661323547,
+      "sampling/importance_sampling_ratio/min": 0.4607382118701935,
+      "sampling/sampling_logp_difference/max": 0.7223583459854126,
+      "sampling/sampling_logp_difference/mean": 0.09794078767299652,
+      "step": 183,
+      "step_time": 10.856938010001613
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7261061072349548,
+      "epoch": 0.00184,
+      "grad_norm": 5.3864270739723e-05,
+      "kl": 0.8039156263694167,
+      "learning_rate": 7.999991890272441e-06,
+      "loss": 0.0,
+      "step": 184,
+      "step_time": 6.554877020000276
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.4375,
+      "completions/mean_terminated_length": 2.4375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9915499612689018,
+      "epoch": 0.00185,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0001638139656279236,
+      "kl": 0.3966022306121886,
+      "learning_rate": 7.999991780311339e-06,
+      "loss": 0.0,
+      "num_tokens": 4436283.0,
+      "reward": 1.3875000476837158,
+      "reward_std": 0.504016101360321,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.4375,
+      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.1218667030334473,
+      "sampling/importance_sampling_ratio/mean": 0.8376058340072632,
+      "sampling/importance_sampling_ratio/min": 0.5223787426948547,
+      "sampling/sampling_logp_difference/max": 0.4979139566421509,
+      "sampling/sampling_logp_difference/mean": 0.09753336012363434,
+      "step": 185,
+      "step_time": 10.397103238996351
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9944529607892036,
+      "epoch": 0.00186,
+      "grad_norm": 0.0002037741505773738,
+      "kl": 0.41308427532203496,
+      "learning_rate": 7.999991669609758e-06,
+      "loss": 0.0,
+      "step": 186,
+      "step_time": 6.079125772997941
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.6875,
+      "completions/mean_terminated_length": 2.6875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8574084043502808,
+      "epoch": 0.00187,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.410769856302068e-05,
+      "kl": 0.5652572922408581,
+      "learning_rate": 7.999991558167702e-06,
+      "loss": 0.0,
+      "num_tokens": 4486596.0,
+      "reward": 2.137500047683716,
+      "reward_std": 1.1482806205749512,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.6875,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.2496119737625122,
+      "sampling/importance_sampling_ratio/mean": 0.8870674967765808,
+      "sampling/importance_sampling_ratio/min": 0.5834121108055115,
+      "sampling/sampling_logp_difference/max": 0.5235366821289062,
+      "sampling/sampling_logp_difference/mean": 0.0739859938621521,
+      "step": 187,
+      "step_time": 11.321346915003232
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8521995544433594,
+      "epoch": 0.00188,
+      "grad_norm": 8.741193596506491e-05,
+      "kl": 0.5736883021891117,
+      "learning_rate": 7.999991445985168e-06,
+      "loss": 0.0,
+      "step": 188,
+      "step_time": 6.144208559999242
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5,
+      "completions/mean_terminated_length": 2.5,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.0526561886072159,
+      "epoch": 0.00189,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.001028874539770186,
+      "kl": 1.272516742348671,
+      "learning_rate": 7.999991333062156e-06,
+      "loss": -0.0,
+      "num_tokens": 4536529.0,
+      "reward": 1.9500000476837158,
+      "reward_std": 1.1639753580093384,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5,
+      "rewards/probe_completion_length/std": 0.5080004930496216,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.0940581560134888,
+      "sampling/importance_sampling_ratio/mean": 0.7959468364715576,
+      "sampling/importance_sampling_ratio/min": 0.2429058998823166,
+      "sampling/sampling_logp_difference/max": 1.2627379894256592,
+      "sampling/sampling_logp_difference/mean": 0.12562313675880432,
+      "step": 189,
+      "step_time": 10.402327937001246
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.065488561987877,
+      "epoch": 0.0019,
+      "grad_norm": 0.001191383576951921,
+      "kl": 1.275945134460926,
+      "learning_rate": 7.999991219398665e-06,
+      "loss": -0.0,
+      "step": 190,
+      "step_time": 6.978599280002527
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.75,
+      "completions/mean_terminated_length": 2.75,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6991342566907406,
+      "epoch": 0.00191,
+      "frac_reward_zero_std": 0.5,
+      "grad_norm": 0.003211932023987174,
+      "kl": 1.2787420004606247,
+      "learning_rate": 7.999991104994699e-06,
+      "loss": -0.0,
+      "num_tokens": 4583658.0,
+      "reward": 2.481250047683716,
+      "reward_std": 1.217728614807129,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.75,
+      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.625,
+      "rewards/probe_shaping_dominance/std": 0.49186936020851135,
+      "rewards/probe_terminal_raw/mean": 0.375,
+      "rewards/probe_terminal_raw/std": 0.49186936020851135,
+      "rewards/rollout_reward_func/mean": -0.25,
+      "rewards/rollout_reward_func/std": 0.9837387204170227,
+      "sampling/importance_sampling_ratio/max": 2.2873668670654297,
+      "sampling/importance_sampling_ratio/mean": 0.9257216453552246,
+      "sampling/importance_sampling_ratio/min": 0.3881727457046509,
+      "sampling/sampling_logp_difference/max": 0.6430342197418213,
+      "sampling/sampling_logp_difference/mean": 0.07942712306976318,
+      "step": 191,
+      "step_time": 10.453309778997209
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7310865595936775,
+      "epoch": 0.00192,
+      "grad_norm": 0.0032008851412683725,
+      "kl": 1.2862421348690987,
+      "learning_rate": 7.999990989850255e-06,
+      "loss": -0.0,
+      "step": 192,
+      "step_time": 6.171673715001816
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.59375,
+      "completions/mean_terminated_length": 2.59375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.0412884652614594,
+      "epoch": 0.00193,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 8.30803910503164e-05,
+      "kl": 0.630632733926177,
+      "learning_rate": 7.999990873965335e-06,
+      "loss": 0.0,
+      "num_tokens": 4627596.0,
+      "reward": 2.043750047683716,
+      "reward_std": 1.201058030128479,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.59375,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.054891586303711,
+      "sampling/importance_sampling_ratio/mean": 0.810993492603302,
+      "sampling/importance_sampling_ratio/min": 0.41359296441078186,
+      "sampling/sampling_logp_difference/max": 0.6084898710250854,
+      "sampling/sampling_logp_difference/mean": 0.10015879571437836,
+      "step": 193,
+      "step_time": 10.796586937998654
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.0507173761725426,
+      "epoch": 0.00194,
+      "grad_norm": 7.215689402073622e-05,
+      "kl": 0.6144167594611645,
+      "learning_rate": 7.999990757339936e-06,
+      "loss": 0.0,
+      "step": 194,
+      "step_time": 6.40111261399943
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.75,
+      "completions/mean_terminated_length": 2.75,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8254596032202244,
+      "epoch": 0.00195,
+      "frac_reward_zero_std": 0.5,
+      "grad_norm": 0.004149310756474733,
+      "kl": 0.899990651756525,
+      "learning_rate": 7.99999063997406e-06,
+      "loss": -0.0,
+      "num_tokens": 4673616.0,
+      "reward": 2.387500047683716,
+      "reward_std": 1.1341474056243896,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.75,
+      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 0.6875,
+      "rewards/probe_shaping_dominance/std": 0.4709290862083435,
+      "rewards/probe_terminal_raw/mean": 0.3125,
+      "rewards/probe_terminal_raw/std": 0.4709290862083435,
+      "rewards/rollout_reward_func/mean": -0.375,
+      "rewards/rollout_reward_func/std": 0.941858172416687,
+      "sampling/importance_sampling_ratio/max": 2.481459617614746,
+      "sampling/importance_sampling_ratio/mean": 0.912665843963623,
+      "sampling/importance_sampling_ratio/min": 0.422697514295578,
+      "sampling/sampling_logp_difference/max": 1.0106472969055176,
+      "sampling/sampling_logp_difference/mean": 0.10286978632211685,
+      "step": 195,
+      "step_time": 10.227664049005398
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8312117978930473,
+      "epoch": 0.00196,
+      "grad_norm": 0.003811546601355076,
+      "kl": 0.9109130464494228,
+      "learning_rate": 7.999990521867708e-06,
+      "loss": -0.0,
+      "step": 196,
+      "step_time": 5.968241761005629
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6872108224779367,
+      "epoch": 0.00197,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.004832283593714237,
+      "kl": 1.1143165528774261,
+      "learning_rate": 7.999990403020877e-06,
+      "loss": -0.0,
+      "num_tokens": 4718553.0,
+      "reward": 2.981250047683716,
+      "reward_std": 1.2309025526046753,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.78125,
+      "rewards/probe_completion_length/std": 0.420013427734375,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 0.40625,
+      "rewards/probe_shaping_dominance/std": 0.49899089336395264,
+      "rewards/probe_terminal_raw/mean": 0.59375,
+      "rewards/probe_terminal_raw/std": 0.49899089336395264,
+      "rewards/rollout_reward_func/mean": 0.1875,
+      "rewards/rollout_reward_func/std": 0.9979817867279053,
+      "sampling/importance_sampling_ratio/max": 2.1045405864715576,
+      "sampling/importance_sampling_ratio/mean": 0.91827791929245,
+      "sampling/importance_sampling_ratio/min": 0.3359697759151459,
+      "sampling/sampling_logp_difference/max": 0.6300592422485352,
+      "sampling/sampling_logp_difference/mean": 0.08973179757595062,
+      "step": 197,
+      "step_time": 10.412749872004497
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.012500000186264515,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.6796288043260574,
+      "epoch": 0.00198,
+      "grad_norm": 0.003753152908757329,
+      "kl": 1.1087056770920753,
+      "learning_rate": 7.99999028343357e-06,
+      "loss": -0.0,
+      "step": 198,
+      "step_time": 6.329036179002287
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.28125,
+      "completions/mean_terminated_length": 2.28125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.141590729355812,
+      "epoch": 0.00199,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00013130051956977695,
+      "kl": 0.556044090539217,
+      "learning_rate": 7.999990163105786e-06,
+      "loss": 0.0,
+      "num_tokens": 4768882.0,
+      "reward": 1.2625000476837158,
+      "reward_std": 0.4709290862083435,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.28125,
+      "rewards/probe_completion_length/std": 0.45680341124534607,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.2970432043075562,
+      "sampling/importance_sampling_ratio/mean": 0.8648799657821655,
+      "sampling/importance_sampling_ratio/min": 0.4939647912979126,
+      "sampling/sampling_logp_difference/max": 0.5484317541122437,
+      "sampling/sampling_logp_difference/mean": 0.0925663411617279,
+      "step": 199,
+      "step_time": 10.656357334992208
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.1397154033184052,
+      "epoch": 0.002,
+      "grad_norm": 0.0001412595302099362,
+      "kl": 0.5653671575710177,
+      "learning_rate": 7.999990042037526e-06,
+      "loss": 0.0,
+      "step": 200,
+      "step_time": 6.58979152199754
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7849503941833973,
+      "epoch": 0.00201,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 7.498179911635816e-05,
+      "kl": 0.4783276170492172,
+      "learning_rate": 7.999989920228787e-06,
+      "loss": 0.0,
+      "num_tokens": 4821237.0,
+      "reward": 1.6375000476837158,
+      "reward_std": 0.4709290862083435,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.65625,
+      "rewards/probe_completion_length/std": 0.4825586974620819,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.0180549621582031,
+      "sampling/importance_sampling_ratio/mean": 0.8009990453720093,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 1.3470954895019531,
+      "sampling/sampling_logp_difference/mean": 0.10532090812921524,
+      "step": 201,
+      "step_time": 11.064135296004679
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7881440408527851,
+      "epoch": 0.00202,
+      "grad_norm": 6.508245132863522e-05,
+      "kl": 0.47384951170533895,
+      "learning_rate": 7.999989797679573e-06,
+      "loss": 0.0,
+      "step": 202,
+      "step_time": 6.172993056996347
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.46875,
+      "completions/mean_terminated_length": 2.46875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8023708537220955,
+      "epoch": 0.00203,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.408317130990326e-05,
+      "kl": 0.8568143546581268,
+      "learning_rate": 7.99998967438988e-06,
+      "loss": 0.0,
+      "num_tokens": 4870596.0,
+      "reward": 1.9187500476837158,
+      "reward_std": 1.256836175918579,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.46875,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.5898252725601196,
+      "sampling/importance_sampling_ratio/mean": 0.9197790026664734,
+      "sampling/importance_sampling_ratio/min": 0.5250067710876465,
+      "sampling/sampling_logp_difference/max": 0.5478343963623047,
+      "sampling/sampling_logp_difference/mean": 0.08068203181028366,
+      "step": 203,
+      "step_time": 11.039577664003446
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.79852594435215,
+      "epoch": 0.00204,
+      "grad_norm": 6.818693509558216e-05,
+      "kl": 0.853181466460228,
+      "learning_rate": 7.999989550359713e-06,
+      "loss": 0.0,
+      "step": 204,
+      "step_time": 6.251118393007346
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 6.0,
+      "completions/max_terminated_length": 6.0,
+      "completions/mean_length": 2.8125,
+      "completions/mean_terminated_length": 2.8125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7093328069895506,
+      "epoch": 0.00205,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 4.685792373493314e-05,
+      "kl": 0.4763212539255619,
+      "learning_rate": 7.999989425589066e-06,
+      "loss": 0.0,
+      "num_tokens": 4913584.0,
+      "reward": 2.8874998092651367,
+      "reward_std": 1.9827888011932373,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.9375,
+      "rewards/probe_completion_length/std": 1.3663583993911743,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5,
+      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
+      "rewards/probe_terminal_raw/mean": 0.5,
+      "rewards/probe_terminal_raw/std": 0.5080004930496216,
+      "rewards/rollout_reward_func/mean": 0.0,
+      "rewards/rollout_reward_func/std": 1.0160009860992432,
+      "sampling/importance_sampling_ratio/max": 1.5179897546768188,
+      "sampling/importance_sampling_ratio/mean": 0.88746577501297,
+      "sampling/importance_sampling_ratio/min": 0.3754726052284241,
+      "sampling/sampling_logp_difference/max": 0.4877147078514099,
+      "sampling/sampling_logp_difference/mean": 0.07970255613327026,
+      "step": 205,
+      "step_time": 10.986664464006026
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7166115269064903,
+      "epoch": 0.00206,
+      "grad_norm": 4.5046446757623926e-05,
+      "kl": 0.4816902196034789,
+      "learning_rate": 7.999989300077943e-06,
+      "loss": 0.0,
+      "step": 206,
+      "step_time": 6.086353101003624
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8723455891013145,
+      "epoch": 0.00207,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0012270144652575254,
+      "kl": 0.8632536679506302,
+      "learning_rate": 7.999989173826344e-06,
+      "loss": -0.0,
+      "num_tokens": 4958457.0,
+      "reward": 2.043750047683716,
+      "reward_std": 1.1175830364227295,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.65625,
+      "rewards/probe_completion_length/std": 0.4825586974620819,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.7422419786453247,
+      "sampling/importance_sampling_ratio/mean": 0.9141074419021606,
+      "sampling/importance_sampling_ratio/min": 0.48550355434417725,
+      "sampling/sampling_logp_difference/max": 0.6161445379257202,
+      "sampling/sampling_logp_difference/mean": 0.10230092704296112,
+      "step": 207,
+      "step_time": 11.013585381002486
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8756029158830643,
+      "epoch": 0.00208,
+      "grad_norm": 0.0018850122578442097,
+      "kl": 0.8681657090783119,
+      "learning_rate": 7.999989046834267e-06,
+      "loss": -0.0,
+      "step": 208,
+      "step_time": 6.54996490999838
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.40625,
+      "completions/mean_terminated_length": 2.40625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8442556709051132,
+      "epoch": 0.00209,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 7.671785715501755e-05,
+      "kl": 0.25706048402935266,
+      "learning_rate": 7.999988919101714e-06,
+      "loss": 0.0,
+      "num_tokens": 5008907.0,
+      "reward": 1.8562500476837158,
+      "reward_std": 1.2790968418121338,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.40625,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.6340703964233398,
+      "sampling/importance_sampling_ratio/mean": 0.915661633014679,
+      "sampling/importance_sampling_ratio/min": 0.49133455753326416,
+      "sampling/sampling_logp_difference/max": 0.6653378009796143,
+      "sampling/sampling_logp_difference/mean": 0.08885866403579712,
+      "step": 209,
+      "step_time": 10.901209290004772
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8593077808618546,
+      "epoch": 0.0021,
+      "grad_norm": 7.776657730573788e-05,
+      "kl": 0.25858927180524915,
+      "learning_rate": 7.999988790628685e-06,
+      "loss": 0.0,
+      "step": 210,
+      "step_time": 6.60464191400024
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.46875,
+      "completions/mean_terminated_length": 2.46875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9752853065729141,
+      "epoch": 0.00211,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.485899939434603e-05,
+      "kl": 0.8968196660280228,
+      "learning_rate": 7.999988661415179e-06,
+      "loss": 0.0,
+      "num_tokens": 5058874.0,
+      "reward": 1.9187500476837158,
+      "reward_std": 1.256836175918579,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.46875,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.2495969533920288,
+      "sampling/importance_sampling_ratio/mean": 0.8685529828071594,
+      "sampling/importance_sampling_ratio/min": 0.2667645215988159,
+      "sampling/sampling_logp_difference/max": 0.712754487991333,
+      "sampling/sampling_logp_difference/mean": 0.09404490143060684,
+      "step": 211,
+      "step_time": 10.649947939000413
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9733599200844765,
+      "epoch": 0.00212,
+      "grad_norm": 0.00013666889572050422,
+      "kl": 0.9042137414216995,
+      "learning_rate": 7.999988531461196e-06,
+      "loss": 0.0,
+      "step": 212,
+      "step_time": 6.069883871998172
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 3.0,
+      "completions/mean_terminated_length": 3.0,
+      "completions/min_length": 3.0,
+      "completions/min_terminated_length": 3.0,
+      "entropy": 0.46625126898288727,
+      "epoch": 0.00213,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0016473316354677081,
+      "kl": 1.0760606080293655,
+      "learning_rate": 7.999988400766736e-06,
+      "loss": -0.0001,
+      "num_tokens": 5101108.0,
+      "reward": 2.825000047683716,
+      "reward_std": 1.008032202720642,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 2.0,
+      "rewards/probe_completion_length/std": 0.0,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5625,
+      "rewards/probe_shaping_dominance/std": 0.504016101360321,
+      "rewards/probe_terminal_raw/mean": 0.4375,
+      "rewards/probe_terminal_raw/std": 0.504016101360321,
+      "rewards/rollout_reward_func/mean": -0.125,
+      "rewards/rollout_reward_func/std": 1.008032202720642,
+      "sampling/importance_sampling_ratio/max": 1.167366862297058,
+      "sampling/importance_sampling_ratio/mean": 0.8593387603759766,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 1.4293286800384521,
+      "sampling/sampling_logp_difference/mean": 0.08223744481801987,
+      "step": 213,
+      "step_time": 10.10010305000469
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.4646361842751503,
+      "epoch": 0.00214,
+      "grad_norm": 0.0016178693622350693,
+      "kl": 1.0777288749814034,
+      "learning_rate": 7.999988269331798e-06,
+      "loss": -0.0001,
+      "step": 214,
+      "step_time": 6.137445069991372
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7770780473947525,
+      "epoch": 0.00215,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 5.397374843596481e-05,
+      "kl": 0.5911668723565526,
+      "learning_rate": 7.999988137156384e-06,
+      "loss": 0.0,
+      "num_tokens": 5150061.0,
+      "reward": 2.137500047683716,
+      "reward_std": 1.1482806205749512,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.0535258054733276,
+      "sampling/importance_sampling_ratio/mean": 0.8605695366859436,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 0.7591410875320435,
+      "sampling/sampling_logp_difference/mean": 0.06761947274208069,
+      "step": 215,
+      "step_time": 11.162840143999347
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7721757888793945,
+      "epoch": 0.00216,
+      "grad_norm": 4.8849575250642374e-05,
+      "kl": 0.5854438902752008,
+      "learning_rate": 7.999988004240496e-06,
+      "loss": 0.0,
+      "step": 216,
+      "step_time": 6.620968937997532
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 6.0,
+      "completions/max_terminated_length": 6.0,
+      "completions/mean_length": 2.71875,
+      "completions/mean_terminated_length": 2.71875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8704147189855576,
+      "epoch": 0.00217,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0052342345006763935,
+      "kl": 1.4794844314455986,
+      "learning_rate": 7.99998787058413e-06,
+      "loss": -0.0002,
+      "num_tokens": 5196138.0,
+      "reward": 2.4499998092651367,
+      "reward_std": 1.4142134189605713,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.6875,
+      "rewards/probe_completion_length/std": 0.6444552540779114,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 0.625,
+      "rewards/probe_shaping_dominance/std": 0.49186936020851135,
+      "rewards/probe_terminal_raw/mean": 0.375,
+      "rewards/probe_terminal_raw/std": 0.49186936020851135,
+      "rewards/rollout_reward_func/mean": -0.25,
+      "rewards/rollout_reward_func/std": 0.9837387204170227,
+      "sampling/importance_sampling_ratio/max": 2.6132655143737793,
+      "sampling/importance_sampling_ratio/mean": 0.9866353273391724,
+      "sampling/importance_sampling_ratio/min": 0.4762972295284271,
+      "sampling/sampling_logp_difference/max": 1.020547866821289,
+      "sampling/sampling_logp_difference/mean": 0.10512237250804901,
+      "step": 217,
+      "step_time": 10.2226877480025
+    },
+    {
+      "clip_ratio/high_max": 0.02500000037252903,
+      "clip_ratio/high_mean": 0.012500000186264515,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.8628628179430962,
+      "epoch": 0.00218,
+      "grad_norm": 0.0029307743534445763,
+      "kl": 1.4838937148451805,
+      "learning_rate": 7.999987736187286e-06,
+      "loss": -0.0002,
+      "step": 218,
+      "step_time": 6.426789073000691
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.46875,
+      "completions/mean_terminated_length": 2.46875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.0259061381220818,
+      "epoch": 0.00219,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00010930380813078955,
+      "kl": 0.5639023052062839,
+      "learning_rate": 7.999987601049967e-06,
+      "loss": 0.0,
+      "num_tokens": 5245170.0,
+      "reward": 1.4500000476837158,
+      "reward_std": 0.5080004930496216,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.46875,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.3624612092971802,
+      "sampling/importance_sampling_ratio/mean": 0.844344437122345,
+      "sampling/importance_sampling_ratio/min": 0.4004921317100525,
+      "sampling/sampling_logp_difference/max": 0.6055986285209656,
+      "sampling/sampling_logp_difference/mean": 0.11378584802150726,
+      "step": 219,
+      "step_time": 10.458324788000027
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.0160942673683167,
+      "epoch": 0.0022,
+      "grad_norm": 0.00012425951717887074,
+      "kl": 0.5864044599002227,
+      "learning_rate": 7.99998746517217e-06,
+      "loss": 0.0,
+      "step": 220,
+      "step_time": 6.467661736995069
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.6875,
+      "completions/mean_terminated_length": 2.6875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7942183464765549,
+      "epoch": 0.00221,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00020853157911915332,
+      "kl": 1.2455565556883812,
+      "learning_rate": 7.999987328553901e-06,
+      "loss": 0.0,
+      "num_tokens": 5291449.0,
+      "reward": 2.637500047683716,
+      "reward_std": 1.3781123161315918,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.6875,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5,
+      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
+      "rewards/probe_terminal_raw/mean": 0.5,
+      "rewards/probe_terminal_raw/std": 0.5080004930496216,
+      "rewards/rollout_reward_func/mean": 0.0,
+      "rewards/rollout_reward_func/std": 1.0160009860992432,
+      "sampling/importance_sampling_ratio/max": 1.0300960540771484,
+      "sampling/importance_sampling_ratio/mean": 0.8702289462089539,
+      "sampling/importance_sampling_ratio/min": 0.3828960061073303,
+      "sampling/sampling_logp_difference/max": 0.6364970803260803,
+      "sampling/sampling_logp_difference/mean": 0.06992574036121368,
+      "step": 221,
+      "step_time": 10.272284876999038
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7945832386612892,
+      "epoch": 0.00222,
+      "grad_norm": 0.0002857858780771494,
+      "kl": 1.2918191030621529,
+      "learning_rate": 7.999987191195152e-06,
+      "loss": 0.0,
+      "step": 222,
+      "step_time": 6.413851582001371
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 8.0,
+      "completions/max_terminated_length": 8.0,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9415403008460999,
+      "epoch": 0.00223,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0001786091597750783,
+      "kl": 0.562560040736571,
+      "learning_rate": 7.999987053095927e-06,
+      "loss": 0.0,
+      "num_tokens": 5339889.0,
+      "reward": 1.6687500476837158,
+      "reward_std": 1.419549584388733,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.71875,
+      "rewards/probe_completion_length/std": 1.419549584388733,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.315468192100525,
+      "sampling/importance_sampling_ratio/mean": 0.8730629682540894,
+      "sampling/importance_sampling_ratio/min": 0.31209444999694824,
+      "sampling/sampling_logp_difference/max": 0.8552637100219727,
+      "sampling/sampling_logp_difference/mean": 0.11184488236904144,
+      "step": 223,
+      "step_time": 10.758617721003247
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9381839409470558,
+      "epoch": 0.00224,
+      "grad_norm": 0.0001709387288428843,
+      "kl": 0.5639883282128721,
+      "learning_rate": 7.999986914256228e-06,
+      "loss": 0.0,
+      "step": 224,
+      "step_time": 6.104447362009523
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.59375,
+      "completions/mean_terminated_length": 2.59375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9080767035484314,
+      "epoch": 0.00225,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.229980499483645e-05,
+      "kl": 0.5537053542211652,
+      "learning_rate": 7.99998677467605e-06,
+      "loss": 0.0,
+      "num_tokens": 5386034.0,
+      "reward": 2.043750047683716,
+      "reward_std": 1.201058030128479,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.59375,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.6336270570755005,
+      "sampling/importance_sampling_ratio/mean": 0.8878433704376221,
+      "sampling/importance_sampling_ratio/min": 0.43646904826164246,
+      "sampling/sampling_logp_difference/max": 0.5756676197052002,
+      "sampling/sampling_logp_difference/mean": 0.09019757807254791,
+      "step": 225,
+      "step_time": 10.606267559003754
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9127962365746498,
+      "epoch": 0.00226,
+      "grad_norm": 0.00010429656686028466,
+      "kl": 0.5661568846553564,
+      "learning_rate": 7.999986634355396e-06,
+      "loss": 0.0,
+      "step": 226,
+      "step_time": 6.377675234001799
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7270364388823509,
+      "epoch": 0.00227,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.001110205426812172,
+      "kl": 0.8677254151552916,
+      "learning_rate": 7.999986493294267e-06,
+      "loss": -0.0001,
+      "num_tokens": 5433028.0,
+      "reward": 2.012500047683716,
+      "reward_std": 1.1341474056243896,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.4258241653442383,
+      "sampling/importance_sampling_ratio/mean": 0.8710789680480957,
+      "sampling/importance_sampling_ratio/min": 0.4317016899585724,
+      "sampling/sampling_logp_difference/max": 0.5112519264221191,
+      "sampling/sampling_logp_difference/mean": 0.07224301993846893,
+      "step": 227,
+      "step_time": 10.239040989999921
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.012500000186264515,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.7276543751358986,
+      "epoch": 0.00228,
+      "grad_norm": 0.0005440651439130306,
+      "kl": 0.8701613224111497,
+      "learning_rate": 7.99998635149266e-06,
+      "loss": -0.0001,
+      "step": 228,
+      "step_time": 6.392367521002598
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.3125,
+      "completions/mean_terminated_length": 2.3125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8749295771121979,
+      "epoch": 0.00229,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 5.688122837455012e-05,
+      "kl": 0.8051012456417084,
+      "learning_rate": 7.99998620895058e-06,
+      "loss": 0.0,
+      "num_tokens": 5482999.0,
+      "reward": 1.7625000476837158,
+      "reward_std": 1.3060035705566406,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.3125,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.0076409578323364,
+      "sampling/importance_sampling_ratio/mean": 0.8914170265197754,
+      "sampling/importance_sampling_ratio/min": 0.4969307780265808,
+      "sampling/sampling_logp_difference/max": 0.6112143993377686,
+      "sampling/sampling_logp_difference/mean": 0.061925433576107025,
+      "step": 229,
+      "step_time": 10.120951806002267
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8805720470845699,
+      "epoch": 0.0023,
+      "grad_norm": 7.190956239355728e-05,
+      "kl": 0.8146018832921982,
+      "learning_rate": 7.999986065668021e-06,
+      "loss": 0.0,
+      "step": 230,
+      "step_time": 6.449205146993336
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.75,
+      "completions/mean_terminated_length": 2.75,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7620595321059227,
+      "epoch": 0.00231,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0013680045958608389,
+      "kl": 0.9012337774038315,
+      "learning_rate": 7.999985921644988e-06,
+      "loss": -0.0,
+      "num_tokens": 5534461.0,
+      "reward": 2.575000047683716,
+      "reward_std": 1.2115039825439453,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.75,
+      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 0.59375,
+      "rewards/probe_shaping_dominance/std": 0.49899089336395264,
+      "rewards/probe_terminal_raw/mean": 0.40625,
+      "rewards/probe_terminal_raw/std": 0.49899089336395264,
+      "rewards/rollout_reward_func/mean": -0.1875,
+      "rewards/rollout_reward_func/std": 0.9979817867279053,
+      "sampling/importance_sampling_ratio/max": 1.1274982690811157,
+      "sampling/importance_sampling_ratio/mean": 0.9018195867538452,
+      "sampling/importance_sampling_ratio/min": 0.5811653733253479,
+      "sampling/sampling_logp_difference/max": 0.36220109462738037,
+      "sampling/sampling_logp_difference/mean": 0.06346160173416138,
+      "step": 231,
+      "step_time": 10.549801017998107
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7599871978163719,
+      "epoch": 0.00232,
+      "grad_norm": 0.0014168447814881802,
+      "kl": 0.8976461552083492,
+      "learning_rate": 7.999985776881479e-06,
+      "loss": -0.0,
+      "step": 232,
+      "step_time": 6.024789627001155
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.8125,
+      "completions/mean_terminated_length": 2.8125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.5355990715324879,
+      "epoch": 0.00233,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0005481232656165957,
+      "kl": 1.2369797490537167,
+      "learning_rate": 7.999985631377493e-06,
+      "loss": 0.0,
+      "num_tokens": 5581452.0,
+      "reward": 2.356250047683716,
+      "reward_std": 0.9791166186332703,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.8125,
+      "rewards/probe_completion_length/std": 0.3965577781200409,
+      "rewards/probe_invalid_count/mean": 0.09375,
+      "rewards/probe_invalid_count/std": 0.2961445748806,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 2.715548276901245,
+      "sampling/importance_sampling_ratio/mean": 0.9138860702514648,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 0.8808169364929199,
+      "sampling/sampling_logp_difference/mean": 0.07896264642477036,
+      "step": 233,
+      "step_time": 10.495318596003926
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.5403720252215862,
+      "epoch": 0.00234,
+      "grad_norm": 0.00044476837501861155,
+      "kl": 1.1751796454191208,
+      "learning_rate": 7.99998548513303e-06,
+      "loss": 0.0,
+      "step": 234,
+      "step_time": 6.561843382994994
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 8.0,
+      "completions/max_terminated_length": 8.0,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8650138974189758,
+      "epoch": 0.00235,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.707937427563593e-05,
+      "kl": 0.4425021205097437,
+      "learning_rate": 7.999985338148094e-06,
+      "loss": 0.0,
+      "num_tokens": 5630802.0,
+      "reward": 2.106250047683716,
+      "reward_std": 1.6482517719268799,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.65625,
+      "rewards/probe_completion_length/std": 1.260040283203125,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.0173451900482178,
+      "sampling/importance_sampling_ratio/mean": 0.832526683807373,
+      "sampling/importance_sampling_ratio/min": 0.4567889869213104,
+      "sampling/sampling_logp_difference/max": 0.6054009199142456,
+      "sampling/sampling_logp_difference/mean": 0.08154759556055069,
+      "step": 235,
+      "step_time": 10.582841627998278
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8746358342468739,
+      "epoch": 0.00236,
+      "grad_norm": 6.966207001823932e-05,
+      "kl": 0.44809109810739756,
+      "learning_rate": 7.99998519042268e-06,
+      "loss": 0.0,
+      "step": 236,
+      "step_time": 6.058184577999782
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.71875,
+      "completions/mean_terminated_length": 2.71875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7308623269200325,
+      "epoch": 0.00237,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.001439909916371107,
+      "kl": 1.1790881492197514,
+      "learning_rate": 7.99998504195679e-06,
+      "loss": -0.0001,
+      "num_tokens": 5679602.0,
+      "reward": 2.606250047683716,
+      "reward_std": 1.334634780883789,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.71875,
+      "rewards/probe_completion_length/std": 0.45680341124534607,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.53125,
+      "rewards/probe_shaping_dominance/std": 0.507007360458374,
+      "rewards/probe_terminal_raw/mean": 0.46875,
+      "rewards/probe_terminal_raw/std": 0.507007360458374,
+      "rewards/rollout_reward_func/mean": -0.0625,
+      "rewards/rollout_reward_func/std": 1.014014720916748,
+      "sampling/importance_sampling_ratio/max": 1.0316462516784668,
+      "sampling/importance_sampling_ratio/mean": 0.8544725179672241,
+      "sampling/importance_sampling_ratio/min": 0.35368672013282776,
+      "sampling/sampling_logp_difference/max": 0.954663097858429,
+      "sampling/sampling_logp_difference/mean": 0.07331576198339462,
+      "step": 237,
+      "step_time": 10.655074249996687
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7382760606706142,
+      "epoch": 0.00238,
+      "grad_norm": 0.0011140814749523997,
+      "kl": 1.187925398349762,
+      "learning_rate": 7.999984892750425e-06,
+      "loss": -0.0001,
+      "step": 238,
+      "step_time": 6.463060008994944
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.75,
+      "completions/mean_terminated_length": 2.75,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.703517246991396,
+      "epoch": 0.00239,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00014675941201858222,
+      "kl": 0.36108433303888887,
+      "learning_rate": 7.999984742803586e-06,
+      "loss": 0.0,
+      "num_tokens": 5727504.0,
+      "reward": 2.200000047683716,
+      "reward_std": 1.1071614027023315,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.75,
+      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.7017192840576172,
+      "sampling/importance_sampling_ratio/mean": 0.8461993336677551,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 1.3302104473114014,
+      "sampling/sampling_logp_difference/mean": 0.08808484673500061,
+      "step": 239,
+      "step_time": 10.698553490001359
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7034162953495979,
+      "epoch": 0.0024,
+      "grad_norm": 0.00018220754282083362,
+      "kl": 0.3764014449552633,
+      "learning_rate": 7.999984592116268e-06,
+      "loss": 0.0,
+      "step": 240,
+      "step_time": 6.1949155629990855
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.7459230236709118,
+      "epoch": 0.00241,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.001529884641058743,
+      "kl": 1.4907125178724527,
+      "learning_rate": 7.999984440688477e-06,
+      "loss": -0.0,
+      "num_tokens": 5777870.0,
+      "reward": 2.481250047683716,
+      "reward_std": 1.3674646615982056,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.65625,
+      "rewards/probe_completion_length/std": 0.4825586974620819,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5625,
+      "rewards/probe_shaping_dominance/std": 0.504016101360321,
+      "rewards/probe_terminal_raw/mean": 0.4375,
+      "rewards/probe_terminal_raw/std": 0.504016101360321,
+      "rewards/rollout_reward_func/mean": -0.125,
+      "rewards/rollout_reward_func/std": 1.008032202720642,
+      "sampling/importance_sampling_ratio/max": 1.2580219507217407,
+      "sampling/importance_sampling_ratio/mean": 0.8700110912322998,
+      "sampling/importance_sampling_ratio/min": 0.267679363489151,
+      "sampling/sampling_logp_difference/max": 1.1994649171829224,
+      "sampling/sampling_logp_difference/mean": 0.09447018057107925,
+      "step": 241,
+      "step_time": 11.393658771001355
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7403427213430405,
+      "epoch": 0.00242,
+      "grad_norm": 0.0016185626154765487,
+      "kl": 1.494697978720069,
+      "learning_rate": 7.999984288520209e-06,
+      "loss": -0.0,
+      "step": 242,
+      "step_time": 6.270897027996398
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.78125,
+      "completions/mean_terminated_length": 2.78125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.744269110262394,
+      "epoch": 0.00243,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0012388663599267602,
+      "kl": 1.5818470269441605,
+      "learning_rate": 7.999984135611465e-06,
+      "loss": 0.0,
+      "num_tokens": 5826374.0,
+      "reward": 2.2312498092651367,
+      "reward_std": 1.0846247673034668,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.84375,
+      "rewards/probe_completion_length/std": 0.5741403102874756,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 0.8125,
+      "rewards/probe_shaping_dominance/std": 0.3965577781200409,
+      "rewards/probe_terminal_raw/mean": 0.1875,
+      "rewards/probe_terminal_raw/std": 0.3965577781200409,
+      "rewards/rollout_reward_func/mean": -0.625,
+      "rewards/rollout_reward_func/std": 0.7931155562400818,
+      "sampling/importance_sampling_ratio/max": 1.8691986799240112,
+      "sampling/importance_sampling_ratio/mean": 0.9386978149414062,
+      "sampling/importance_sampling_ratio/min": 0.31528666615486145,
+      "sampling/sampling_logp_difference/max": 0.8249779343605042,
+      "sampling/sampling_logp_difference/mean": 0.10700443387031555,
+      "step": 243,
+      "step_time": 10.129046364996611
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7368881106376648,
+      "epoch": 0.00244,
+      "grad_norm": 0.005345514975488186,
+      "kl": 1.592999592423439,
+      "learning_rate": 7.999983981962246e-06,
+      "loss": 0.0,
+      "step": 244,
+      "step_time": 6.396396995001851
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.6875,
+      "completions/mean_terminated_length": 2.6875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8525449559092522,
+      "epoch": 0.00245,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0003305396530777216,
+      "kl": 1.2175440937280655,
+      "learning_rate": 7.999983827572551e-06,
+      "loss": 0.0,
+      "num_tokens": 5873940.0,
+      "reward": 2.137500047683716,
+      "reward_std": 1.1482806205749512,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.6875,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.0157396793365479,
+      "sampling/importance_sampling_ratio/mean": 0.8005120754241943,
+      "sampling/importance_sampling_ratio/min": 0.24334916472434998,
+      "sampling/sampling_logp_difference/max": 1.1724286079406738,
+      "sampling/sampling_logp_difference/mean": 0.11140745878219604,
+      "step": 245,
+      "step_time": 10.66775867099932
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8477492108941078,
+      "epoch": 0.00246,
+      "grad_norm": 0.0003489796072244644,
+      "kl": 1.2276836708188057,
+      "learning_rate": 7.999983672442382e-06,
+      "loss": 0.0,
+      "step": 246,
+      "step_time": 5.994220978001977
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.1875,
+      "completions/mean_terminated_length": 2.1875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.1669961959123611,
+      "epoch": 0.00247,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 5.2477411372819915e-05,
+      "kl": 0.13767928373999894,
+      "learning_rate": 7.999983516571737e-06,
+      "loss": 0.0,
+      "num_tokens": 5925543.0,
+      "reward": 1.1375000476837158,
+      "reward_std": 0.3965577781200409,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.1875,
+      "rewards/probe_completion_length/std": 0.3965577781200409,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.14869225025177,
+      "sampling/importance_sampling_ratio/mean": 0.8607889413833618,
+      "sampling/importance_sampling_ratio/min": 0.5642183423042297,
+      "sampling/sampling_logp_difference/max": 0.4303678274154663,
+      "sampling/sampling_logp_difference/mean": 0.09759242087602615,
+      "step": 247,
+      "step_time": 10.611091466998914
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.1723481938242912,
+      "epoch": 0.00248,
+      "grad_norm": 3.973311686422676e-05,
+      "kl": 0.13221220299601555,
+      "learning_rate": 7.999983359960615e-06,
+      "loss": 0.0,
+      "step": 248,
+      "step_time": 6.086148332004086
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.687354326248169,
+      "epoch": 0.00249,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0011157306144014,
+      "kl": 1.323385737836361,
+      "learning_rate": 7.999983202609019e-06,
+      "loss": -0.0001,
+      "num_tokens": 5970333.0,
+      "reward": 2.543750047683716,
+      "reward_std": 1.387952208518982,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.65625,
+      "rewards/probe_completion_length/std": 0.4825586974620819,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.53125,
+      "rewards/probe_shaping_dominance/std": 0.507007360458374,
+      "rewards/probe_terminal_raw/mean": 0.46875,
+      "rewards/probe_terminal_raw/std": 0.507007360458374,
+      "rewards/rollout_reward_func/mean": -0.0625,
+      "rewards/rollout_reward_func/std": 1.014014720916748,
+      "sampling/importance_sampling_ratio/max": 1.8828648328781128,
+      "sampling/importance_sampling_ratio/mean": 0.9965871572494507,
+      "sampling/importance_sampling_ratio/min": 0.3140735924243927,
+      "sampling/sampling_logp_difference/max": 0.9687585830688477,
+      "sampling/sampling_logp_difference/mean": 0.11039839684963226,
+      "step": 249,
+      "step_time": 10.432378074001463
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6846778839826584,
+      "epoch": 0.0025,
+      "grad_norm": 0.0010683604050427675,
+      "kl": 1.3352181240916252,
+      "learning_rate": 7.999983044516948e-06,
+      "loss": -0.0001,
+      "step": 250,
+      "step_time": 5.937322461002623
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.4375,
+      "completions/mean_terminated_length": 2.4375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.156593695282936,
+      "epoch": 0.00251,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00012909414363093674,
+      "kl": 1.1634539794176817,
+      "learning_rate": 7.999982885684401e-06,
+      "loss": 0.0,
+      "num_tokens": 6016418.0,
+      "reward": 1.4500000476837158,
+      "reward_std": 0.5080004930496216,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.4375,
+      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.2055965662002563,
+      "sampling/importance_sampling_ratio/mean": 0.8325642347335815,
+      "sampling/importance_sampling_ratio/min": 0.517051100730896,
+      "sampling/sampling_logp_difference/max": 0.548661470413208,
+      "sampling/sampling_logp_difference/mean": 0.10579308122396469,
+      "step": 251,
+      "step_time": 10.488625730999047
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.1303688883781433,
+      "epoch": 0.00252,
+      "grad_norm": 0.00012144693755544722,
+      "kl": 1.1275506392121315,
+      "learning_rate": 7.99998272611138e-06,
+      "loss": 0.0,
+      "step": 252,
+      "step_time": 5.996650919005333
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.25,
+      "completions/mean_terminated_length": 2.25,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9809732586145401,
+      "epoch": 0.00253,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.001729611773043871,
+      "kl": 1.0025988789275289,
+      "learning_rate": 7.999982565797882e-06,
+      "loss": 0.0,
+      "num_tokens": 6068332.0,
+      "reward": 1.2000000476837158,
+      "reward_std": 0.4399413466453552,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.25,
+      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 2.0689034461975098,
+      "sampling/importance_sampling_ratio/mean": 0.9019052982330322,
+      "sampling/importance_sampling_ratio/min": 0.4535451829433441,
+      "sampling/sampling_logp_difference/max": 0.8029270172119141,
+      "sampling/sampling_logp_difference/mean": 0.10248970985412598,
+      "step": 253,
+      "step_time": 10.398529680995125
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9765767902135849,
+      "epoch": 0.00254,
+      "grad_norm": 0.0014561200514435768,
+      "kl": 0.9909821730107069,
+      "learning_rate": 7.999982404743908e-06,
+      "loss": 0.0,
+      "step": 254,
+      "step_time": 6.541713023001648
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6780846863985062,
+      "epoch": 0.00255,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0011963710421696305,
+      "kl": 1.390501230955124,
+      "learning_rate": 7.999982242949461e-06,
+      "loss": -0.0,
+      "num_tokens": 6117491.0,
+      "reward": 2.387500047683716,
+      "reward_std": 1.2935945987701416,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 0.625,
+      "rewards/probe_shaping_dominance/std": 0.49186936020851135,
+      "rewards/probe_terminal_raw/mean": 0.375,
+      "rewards/probe_terminal_raw/std": 0.49186936020851135,
+      "rewards/rollout_reward_func/mean": -0.25,
+      "rewards/rollout_reward_func/std": 0.9837387204170227,
+      "sampling/importance_sampling_ratio/max": 1.108931303024292,
+      "sampling/importance_sampling_ratio/mean": 0.8845682144165039,
+      "sampling/importance_sampling_ratio/min": 0.626857340335846,
+      "sampling/sampling_logp_difference/max": 0.38432884216308594,
+      "sampling/sampling_logp_difference/mean": 0.06597592681646347,
+      "step": 255,
+      "step_time": 10.823603749999165
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6782475523650646,
+      "epoch": 0.00256,
+      "grad_norm": 0.001205417443998158,
+      "kl": 1.3910483717918396,
+      "learning_rate": 7.999982080414539e-06,
+      "loss": -0.0,
+      "step": 256,
+      "step_time": 6.142732071006321
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.4375,
+      "completions/mean_terminated_length": 2.4375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8866551741957664,
+      "epoch": 0.00257,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00015870347851887345,
+      "kl": 0.3300934328290168,
+      "learning_rate": 7.999981917139141e-06,
+      "loss": 0.0,
+      "num_tokens": 6163105.0,
+      "reward": 1.8875000476837158,
+      "reward_std": 1.2684128284454346,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.4375,
+      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.006065011024475,
+      "sampling/importance_sampling_ratio/mean": 0.8799917101860046,
+      "sampling/importance_sampling_ratio/min": 0.4552786946296692,
+      "sampling/sampling_logp_difference/max": 0.48933935165405273,
+      "sampling/sampling_logp_difference/mean": 0.07220719009637833,
+      "step": 257,
+      "step_time": 10.466877447997831
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8837831430137157,
+      "epoch": 0.00258,
+      "grad_norm": 0.0001448682160116732,
+      "kl": 0.32037485315231606,
+      "learning_rate": 7.999981753123268e-06,
+      "loss": 0.0,
+      "step": 258,
+      "step_time": 6.370652053996309
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.65625,
+      "completions/mean_terminated_length": 2.65625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6461021453142166,
+      "epoch": 0.00259,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.003029422601684928,
+      "kl": 2.338309422135353,
+      "learning_rate": 7.999981588366921e-06,
+      "loss": -0.0001,
+      "num_tokens": 6210069.0,
+      "reward": 2.418750047683716,
+      "reward_std": 1.343667984008789,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.65625,
+      "rewards/probe_completion_length/std": 0.4825586974620819,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.59375,
+      "rewards/probe_shaping_dominance/std": 0.49899089336395264,
+      "rewards/probe_terminal_raw/mean": 0.40625,
+      "rewards/probe_terminal_raw/std": 0.49899089336395264,
+      "rewards/rollout_reward_func/mean": -0.1875,
+      "rewards/rollout_reward_func/std": 0.9979817867279053,
+      "sampling/importance_sampling_ratio/max": 1.1730843782424927,
+      "sampling/importance_sampling_ratio/mean": 0.8798280954360962,
+      "sampling/importance_sampling_ratio/min": 0.41403934359550476,
+      "sampling/sampling_logp_difference/max": 0.829841136932373,
+      "sampling/sampling_logp_difference/mean": 0.07408564537763596,
+      "step": 259,
+      "step_time": 10.038688091004587
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6447968520224094,
+      "epoch": 0.0026,
+      "grad_norm": 0.0027008464094251394,
+      "kl": 2.210945799946785,
+      "learning_rate": 7.999981422870099e-06,
+      "loss": -0.0001,
+      "step": 260,
+      "step_time": 6.00890496500142
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.71875,
+      "completions/mean_terminated_length": 2.71875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.6659299023449421,
+      "epoch": 0.00261,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.075092824175954e-05,
+      "kl": 1.107674665749073,
+      "learning_rate": 7.999981256632802e-06,
+      "loss": 0.0,
+      "num_tokens": 6258892.0,
+      "reward": 2.168750047683716,
+      "reward_std": 1.1283552646636963,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.71875,
+      "rewards/probe_completion_length/std": 0.45680341124534607,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.1683428287506104,
+      "sampling/importance_sampling_ratio/mean": 0.9052141904830933,
+      "sampling/importance_sampling_ratio/min": 0.4793435037136078,
+      "sampling/sampling_logp_difference/max": 0.5965175628662109,
+      "sampling/sampling_logp_difference/mean": 0.05744236335158348,
+      "step": 261,
+      "step_time": 10.571983769008511
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.6678782217204571,
+      "epoch": 0.00262,
+      "grad_norm": 5.548512490349822e-05,
+      "kl": 1.1050898134708405,
+      "learning_rate": 7.999981089655028e-06,
+      "loss": 0.0,
+      "step": 262,
+      "step_time": 6.452769112001988
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5625,
+      "completions/mean_terminated_length": 2.5625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9660641402006149,
+      "epoch": 0.00263,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.0001927059784065932,
+      "kl": 0.8233552714809775,
+      "learning_rate": 7.999980921936782e-06,
+      "loss": 0.0,
+      "num_tokens": 6305938.0,
+      "reward": 1.5437500476837158,
+      "reward_std": 0.498990923166275,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5625,
+      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.3487149477005005,
+      "sampling/importance_sampling_ratio/mean": 0.8366470336914062,
+      "sampling/importance_sampling_ratio/min": 0.27520906925201416,
+      "sampling/sampling_logp_difference/max": 0.9875253438949585,
+      "sampling/sampling_logp_difference/mean": 0.11163996160030365,
+      "step": 263,
+      "step_time": 10.18768118799926
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9702360332012177,
+      "epoch": 0.00264,
+      "grad_norm": 0.00022108059783931822,
+      "kl": 0.8513107509352267,
+      "learning_rate": 7.999980753478058e-06,
+      "loss": 0.0,
+      "step": 264,
+      "step_time": 6.391760288996011
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.6875,
+      "completions/mean_terminated_length": 2.6875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9113809689879417,
+      "epoch": 0.00265,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00035716500133275986,
+      "kl": 1.2488736808300018,
+      "learning_rate": 7.999980584278861e-06,
+      "loss": 0.0,
+      "num_tokens": 6353928.0,
+      "reward": 1.6687500476837158,
+      "reward_std": 0.45680341124534607,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.6875,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 2.03200101852417,
+      "sampling/importance_sampling_ratio/mean": 0.8922456502914429,
+      "sampling/importance_sampling_ratio/min": 0.2198769450187683,
+      "sampling/sampling_logp_difference/max": 1.1652178764343262,
+      "sampling/sampling_logp_difference/mean": 0.13655489683151245,
+      "step": 265,
+      "step_time": 10.298811807006132
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9102443382143974,
+      "epoch": 0.00266,
+      "grad_norm": 0.00027845395379699767,
+      "kl": 1.2205140069127083,
+      "learning_rate": 7.999980414339192e-06,
+      "loss": 0.0,
+      "step": 266,
+      "step_time": 6.431490440001653
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 2017.0,
+      "completions/max_terminated_length": 2017.0,
+      "completions/mean_length": 65.40625,
+      "completions/mean_terminated_length": 65.40625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.0315296947956085,
+      "epoch": 0.00267,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00012318712833803147,
+      "kl": 0.6766951990430243,
+      "learning_rate": 7.999980243659046e-06,
+      "loss": 0.0,
+      "num_tokens": 6404030.0,
+      "reward": 187.01251220703125,
+      "reward_std": 1047.0548095703125,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 186.5625,
+      "rewards/probe_completion_length/std": 1047.1453857421875,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.0069550275802612,
+      "sampling/importance_sampling_ratio/mean": 0.8316231966018677,
+      "sampling/importance_sampling_ratio/min": 0.0,
+      "sampling/sampling_logp_difference/max": 2.4841392040252686,
+      "sampling/sampling_logp_difference/mean": 0.25313299894332886,
+      "step": 267,
+      "step_time": 33.58669508300227
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.0328385680913925,
+      "epoch": 0.00268,
+      "grad_norm": 0.00010500734060769901,
+      "kl": 0.6582283083116636,
+      "learning_rate": 7.999980072238424e-06,
+      "loss": 0.0,
+      "step": 268,
+      "step_time": 11.70861316099763
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.53125,
+      "completions/mean_terminated_length": 2.53125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8391192331910133,
+      "epoch": 0.00269,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00012727642024401575,
+      "kl": 1.1651003416627645,
+      "learning_rate": 7.999979900077329e-06,
+      "loss": 0.0,
+      "num_tokens": 6454454.0,
+      "reward": 2.075000047683716,
+      "reward_std": 1.1845782995224,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.53125,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.09375,
+      "rewards/probe_invalid_count/std": 0.2961445748806,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 2.698387622833252,
+      "sampling/importance_sampling_ratio/mean": 0.9577848315238953,
+      "sampling/importance_sampling_ratio/min": 0.3580544888973236,
+      "sampling/sampling_logp_difference/max": 1.1420834064483643,
+      "sampling/sampling_logp_difference/mean": 0.11074647307395935,
+      "step": 269,
+      "step_time": 10.692912369002443
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8405395671725273,
+      "epoch": 0.0027,
+      "grad_norm": 0.00012962006439920515,
+      "kl": 1.1688371929340065,
+      "learning_rate": 7.99997972717576e-06,
+      "loss": 0.0,
+      "step": 270,
+      "step_time": 6.066991922998568
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5625,
+      "completions/mean_terminated_length": 2.5625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.69916364364326,
+      "epoch": 0.00271,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 2.587308881629724e-05,
+      "kl": 0.6396325055975467,
+      "learning_rate": 7.999979553533716e-06,
+      "loss": 0.0,
+      "num_tokens": 6498469.0,
+      "reward": 2.512500047683716,
+      "reward_std": 1.4797013998031616,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5625,
+      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5,
+      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
+      "rewards/probe_terminal_raw/mean": 0.5,
+      "rewards/probe_terminal_raw/std": 0.5080004930496216,
+      "rewards/rollout_reward_func/mean": 0.0,
+      "rewards/rollout_reward_func/std": 1.0160009860992432,
+      "sampling/importance_sampling_ratio/max": 1.1648913621902466,
+      "sampling/importance_sampling_ratio/mean": 0.9113304615020752,
+      "sampling/importance_sampling_ratio/min": 0.6570224165916443,
+      "sampling/sampling_logp_difference/max": 0.3369510769844055,
+      "sampling/sampling_logp_difference/mean": 0.05420326441526413,
+      "step": 271,
+      "step_time": 10.439027642998553
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.7068110965192318,
+      "epoch": 0.00272,
+      "grad_norm": 2.776672408799641e-05,
+      "kl": 0.6432693933602422,
+      "learning_rate": 7.999979379151197e-06,
+      "loss": 0.0,
+      "step": 272,
+      "step_time": 6.350383502001932
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.96875,
+      "completions/mean_terminated_length": 2.96875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.502995315939188,
+      "epoch": 0.00273,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00011989998165518045,
+      "kl": 1.7036965787410736,
+      "learning_rate": 7.999979204028205e-06,
+      "loss": 0.0,
+      "num_tokens": 6544886.0,
+      "reward": 2.918750047683716,
+      "reward_std": 1.0620847940444946,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.96875,
+      "rewards/probe_completion_length/std": 0.1767766922712326,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.5,
+      "rewards/probe_shaping_dominance/std": 0.5080004930496216,
+      "rewards/probe_terminal_raw/mean": 0.5,
+      "rewards/probe_terminal_raw/std": 0.5080004930496216,
+      "rewards/rollout_reward_func/mean": 0.0,
+      "rewards/rollout_reward_func/std": 1.0160009860992432,
+      "sampling/importance_sampling_ratio/max": 1.664729118347168,
+      "sampling/importance_sampling_ratio/mean": 0.930694580078125,
+      "sampling/importance_sampling_ratio/min": 0.36706072092056274,
+      "sampling/sampling_logp_difference/max": 0.884929895401001,
+      "sampling/sampling_logp_difference/mean": 0.08341725915670395,
+      "step": 273,
+      "step_time": 10.285292349999509
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.5101733300834894,
+      "epoch": 0.00274,
+      "grad_norm": 9.926714847097173e-05,
+      "kl": 1.6980668678879738,
+      "learning_rate": 7.999979028164737e-06,
+      "loss": 0.0,
+      "step": 274,
+      "step_time": 6.408322647002933
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.010416666977107525,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.010416666977107525,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 7.0,
+      "completions/max_terminated_length": 7.0,
+      "completions/mean_length": 2.96875,
+      "completions/mean_terminated_length": 2.96875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.713752806186676,
+      "epoch": 0.00275,
+      "frac_reward_zero_std": 0.5,
+      "grad_norm": 0.0021332697942852974,
+      "kl": 1.2537326663732529,
+      "learning_rate": 7.999978851560795e-06,
+      "loss": -0.0001,
+      "num_tokens": 6591867.0,
+      "reward": 2.5437498092651367,
+      "reward_std": 1.4560080766677856,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 2.03125,
+      "rewards/probe_completion_length/std": 1.1495966911315918,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.71875,
+      "rewards/probe_shaping_dominance/std": 0.45680341124534607,
+      "rewards/probe_terminal_raw/mean": 0.28125,
+      "rewards/probe_terminal_raw/std": 0.45680341124534607,
+      "rewards/rollout_reward_func/mean": -0.4375,
+      "rewards/rollout_reward_func/std": 0.9136068224906921,
+      "sampling/importance_sampling_ratio/max": 1.455705165863037,
+      "sampling/importance_sampling_ratio/mean": 0.8554098010063171,
+      "sampling/importance_sampling_ratio/min": 0.25982654094696045,
+      "sampling/sampling_logp_difference/max": 1.1220810413360596,
+      "sampling/sampling_logp_difference/mean": 0.11255955696105957,
+      "step": 275,
+      "step_time": 10.903306543001236
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.012500000186264515,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.012500000186264515,
+      "entropy": 0.7152140066027641,
+      "epoch": 0.00276,
+      "grad_norm": 0.002171127125620842,
+      "kl": 1.241127960383892,
+      "learning_rate": 7.999978674216379e-06,
+      "loss": -0.0001,
+      "step": 276,
+      "step_time": 6.070132431003003
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.53125,
+      "completions/mean_terminated_length": 2.53125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9556013569235802,
+      "epoch": 0.00277,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.371323252096772e-05,
+      "kl": 0.6075592692941427,
+      "learning_rate": 7.99997849613149e-06,
+      "loss": 0.0,
+      "num_tokens": 6639069.0,
+      "reward": 1.9812500476837158,
+      "reward_std": 1.2309025526046753,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.53125,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.75,
+      "rewards/probe_shaping_dominance/std": 0.4399413466453552,
+      "rewards/probe_terminal_raw/mean": 0.25,
+      "rewards/probe_terminal_raw/std": 0.4399413466453552,
+      "rewards/rollout_reward_func/mean": -0.5,
+      "rewards/rollout_reward_func/std": 0.8798826932907104,
+      "sampling/importance_sampling_ratio/max": 1.5697556734085083,
+      "sampling/importance_sampling_ratio/mean": 0.873795747756958,
+      "sampling/importance_sampling_ratio/min": 0.5771368145942688,
+      "sampling/sampling_logp_difference/max": 0.5684585571289062,
+      "sampling/sampling_logp_difference/mean": 0.08412325382232666,
+      "step": 277,
+      "step_time": 10.163918867001485
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9650540724396706,
+      "epoch": 0.00278,
+      "grad_norm": 6.126017979113385e-05,
+      "kl": 0.601465885527432,
+      "learning_rate": 7.999978317306126e-06,
+      "loss": 0.0,
+      "step": 278,
+      "step_time": 6.426909692010668
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.53125,
+      "completions/mean_terminated_length": 2.53125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.0120504274964333,
+      "epoch": 0.00279,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.000864097208250314,
+      "kl": 1.258582164766267,
+      "learning_rate": 7.999978137740288e-06,
+      "loss": 0.0,
+      "num_tokens": 6688720.0,
+      "reward": 1.5125000476837158,
+      "reward_std": 0.504016101360321,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.53125,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 2.874281406402588,
+      "sampling/importance_sampling_ratio/mean": 0.8551193475723267,
+      "sampling/importance_sampling_ratio/min": 0.2749137282371521,
+      "sampling/sampling_logp_difference/max": 1.2301058769226074,
+      "sampling/sampling_logp_difference/mean": 0.1270916908979416,
+      "step": 279,
+      "step_time": 10.681570811004349
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.014964997768402,
+      "epoch": 0.0028,
+      "grad_norm": 0.0005978773115202785,
+      "kl": 1.168176197214052,
+      "learning_rate": 7.999977957433975e-06,
+      "loss": 0.0,
+      "step": 280,
+      "step_time": 6.049058700999012
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.1875,
+      "completions/mean_terminated_length": 2.1875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.3366922959685326,
+      "epoch": 0.00281,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00011146832548547536,
+      "kl": 0.5858726419537561,
+      "learning_rate": 7.999977776387188e-06,
+      "loss": 0.0,
+      "num_tokens": 6735799.0,
+      "reward": 1.1375000476837158,
+      "reward_std": 0.3965577781200409,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.1875,
+      "rewards/probe_completion_length/std": 0.3965577781200409,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.3003031015396118,
+      "sampling/importance_sampling_ratio/mean": 0.84781813621521,
+      "sampling/importance_sampling_ratio/min": 0.4803212285041809,
+      "sampling/sampling_logp_difference/max": 0.5285500288009644,
+      "sampling/sampling_logp_difference/mean": 0.10937441885471344,
+      "step": 281,
+      "step_time": 11.007205139005237
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.3633311539888382,
+      "epoch": 0.00282,
+      "grad_norm": 0.00011356321920175105,
+      "kl": 0.5938262157142162,
+      "learning_rate": 7.999977594599927e-06,
+      "loss": 0.0,
+      "step": 282,
+      "step_time": 6.034540549997473
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.96875,
+      "completions/mean_terminated_length": 2.96875,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.3489413410425186,
+      "epoch": 0.00283,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0025514166336506605,
+      "kl": 4.161712437868118,
+      "learning_rate": 7.999977412072193e-06,
+      "loss": 0.0,
+      "num_tokens": 6780884.0,
+      "reward": 3.137500047683716,
+      "reward_std": 0.99798184633255,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.96875,
+      "rewards/probe_completion_length/std": 0.1767766922712326,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.40625,
+      "rewards/probe_shaping_dominance/std": 0.49899089336395264,
+      "rewards/probe_terminal_raw/mean": 0.59375,
+      "rewards/probe_terminal_raw/std": 0.49899089336395264,
+      "rewards/rollout_reward_func/mean": 0.1875,
+      "rewards/rollout_reward_func/std": 0.9979817867279053,
+      "sampling/importance_sampling_ratio/max": 2.1546339988708496,
+      "sampling/importance_sampling_ratio/mean": 0.9989504814147949,
+      "sampling/importance_sampling_ratio/min": 0.4079555869102478,
+      "sampling/sampling_logp_difference/max": 0.9860153198242188,
+      "sampling/sampling_logp_difference/mean": 0.051779210567474365,
+      "step": 283,
+      "step_time": 10.310512965006637
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.3527596667408943,
+      "epoch": 0.00284,
+      "grad_norm": 0.0023123540449887514,
+      "kl": 4.159931391477585,
+      "learning_rate": 7.999977228803984e-06,
+      "loss": 0.0,
+      "step": 284,
+      "step_time": 5.892429111005185
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5625,
+      "completions/mean_terminated_length": 2.5625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.0530920773744583,
+      "epoch": 0.00285,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 5.824619802297093e-05,
+      "kl": 0.7229606504552066,
+      "learning_rate": 7.999977044795302e-06,
+      "loss": 0.0,
+      "num_tokens": 6831378.0,
+      "reward": 1.5750000476837158,
+      "reward_std": 0.49186936020851135,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5625,
+      "rewards/probe_completion_length/std": 0.504016101360321,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.3243030309677124,
+      "sampling/importance_sampling_ratio/mean": 0.8309626579284668,
+      "sampling/importance_sampling_ratio/min": 0.4618091583251953,
+      "sampling/sampling_logp_difference/max": 0.5776401162147522,
+      "sampling/sampling_logp_difference/mean": 0.11181661486625671,
+      "step": 285,
+      "step_time": 10.82583777599939
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.0667740106582642,
+      "epoch": 0.00286,
+      "grad_norm": 6.092442708904855e-05,
+      "kl": 0.7223768094554543,
+      "learning_rate": 7.999976860046145e-06,
+      "loss": 0.0,
+      "step": 286,
+      "step_time": 6.030480374996841
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 4.0,
+      "completions/max_terminated_length": 4.0,
+      "completions/mean_length": 2.5625,
+      "completions/mean_terminated_length": 2.5625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.1873224079608917,
+      "epoch": 0.00287,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 0.00046929955715313554,
+      "kl": 0.7589781978167593,
+      "learning_rate": 7.999976674556518e-06,
+      "loss": 0.0,
+      "num_tokens": 6883342.0,
+      "reward": 1.5750000476837158,
+      "reward_std": 0.6599119901657104,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5625,
+      "rewards/probe_completion_length/std": 0.5644009113311768,
+      "rewards/probe_invalid_count/mean": 0.0625,
+      "rewards/probe_invalid_count/std": 0.24593468010425568,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 2.714132070541382,
+      "sampling/importance_sampling_ratio/mean": 0.8575228452682495,
+      "sampling/importance_sampling_ratio/min": 0.27807188034057617,
+      "sampling/sampling_logp_difference/max": 1.1784460544586182,
+      "sampling/sampling_logp_difference/mean": 0.15426218509674072,
+      "step": 287,
+      "step_time": 10.328236994999315
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.205030769109726,
+      "epoch": 0.00288,
+      "grad_norm": 0.0003330877225380391,
+      "kl": 0.7408322375267744,
+      "learning_rate": 7.999976488326414e-06,
+      "loss": 0.0,
+      "step": 288,
+      "step_time": 6.8534642979975615
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.5,
+      "completions/mean_terminated_length": 2.5,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9017446711659431,
+      "epoch": 0.00289,
+      "frac_reward_zero_std": 1.0,
+      "grad_norm": 6.207494152477011e-05,
+      "kl": 0.89900878444314,
+      "learning_rate": 7.999976301355836e-06,
+      "loss": 0.0,
+      "num_tokens": 6929162.0,
+      "reward": 1.4500000476837158,
+      "reward_std": 0.5080004930496216,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.5,
+      "rewards/probe_completion_length/std": 0.5080004930496216,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 1.0,
+      "rewards/probe_shaping_dominance/std": 0.0,
+      "rewards/probe_terminal_raw/mean": 0.0,
+      "rewards/probe_terminal_raw/std": 0.0,
+      "rewards/rollout_reward_func/mean": -1.0,
+      "rewards/rollout_reward_func/std": 0.0,
+      "sampling/importance_sampling_ratio/max": 1.3137609958648682,
+      "sampling/importance_sampling_ratio/mean": 0.8832715153694153,
+      "sampling/importance_sampling_ratio/min": 0.5119587182998657,
+      "sampling/sampling_logp_difference/max": 0.3956003189086914,
+      "sampling/sampling_logp_difference/mean": 0.07285954803228378,
+      "step": 289,
+      "step_time": 10.149409565994574
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9096414819359779,
+      "epoch": 0.0029,
+      "grad_norm": 7.079554052324966e-05,
+      "kl": 0.9083996750414371,
+      "learning_rate": 7.999976113644787e-06,
+      "loss": 0.0,
+      "step": 290,
+      "step_time": 5.997940316992754
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.53125,
+      "completions/mean_terminated_length": 2.53125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.046548217535019,
+      "epoch": 0.00291,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0014521953416988254,
+      "kl": 1.2034405604936182,
+      "learning_rate": 7.999975925193261e-06,
+      "loss": -0.0001,
+      "num_tokens": 6981084.0,
+      "reward": 1.9187500476837158,
+      "reward_std": 1.1773227453231812,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.53125,
+      "rewards/probe_completion_length/std": 0.507007360458374,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 1.1272008419036865,
+      "sampling/importance_sampling_ratio/mean": 0.8361198902130127,
+      "sampling/importance_sampling_ratio/min": 0.44038787484169006,
+      "sampling/sampling_logp_difference/max": 0.7219665050506592,
+      "sampling/sampling_logp_difference/mean": 0.11340387165546417,
+      "step": 291,
+      "step_time": 11.30698675900203
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.0575671941041946,
+      "epoch": 0.00292,
+      "grad_norm": 0.0013669326435774565,
+      "kl": 1.2194756520912051,
+      "learning_rate": 7.999975736001263e-06,
+      "loss": -0.0001,
+      "step": 292,
+      "step_time": 6.158662155998172
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.59375,
+      "completions/mean_terminated_length": 2.59375,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9219824075698853,
+      "epoch": 0.00293,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0009061606251634657,
+      "kl": 1.331733226776123,
+      "learning_rate": 7.999975546068793e-06,
+      "loss": -0.0,
+      "num_tokens": 7031701.0,
+      "reward": 1.7937500476837158,
+      "reward_std": 0.9540871381759644,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.59375,
+      "rewards/probe_completion_length/std": 0.49899089336395264,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.875,
+      "rewards/probe_shaping_dominance/std": 0.33601075410842896,
+      "rewards/probe_terminal_raw/mean": 0.125,
+      "rewards/probe_terminal_raw/std": 0.33601075410842896,
+      "rewards/rollout_reward_func/mean": -0.75,
+      "rewards/rollout_reward_func/std": 0.6720215082168579,
+      "sampling/importance_sampling_ratio/max": 1.3209359645843506,
+      "sampling/importance_sampling_ratio/mean": 0.8001704216003418,
+      "sampling/importance_sampling_ratio/min": 0.15668565034866333,
+      "sampling/sampling_logp_difference/max": 1.3062156438827515,
+      "sampling/sampling_logp_difference/mean": 0.16086944937705994,
+      "step": 293,
+      "step_time": 10.72736765299851
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.9374803677201271,
+      "epoch": 0.00294,
+      "grad_norm": 0.0009243946406058967,
+      "kl": 1.3495024144649506,
+      "learning_rate": 7.999975355395847e-06,
+      "loss": -0.0,
+      "step": 294,
+      "step_time": 6.198022962998948
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.625,
+      "completions/mean_terminated_length": 2.625,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.8686415776610374,
+      "epoch": 0.00295,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.002907541813328862,
+      "kl": 1.5130201950669289,
+      "learning_rate": 7.999975163982429e-06,
+      "loss": -0.0,
+      "num_tokens": 7078086.0,
+      "reward": 2.512500047683716,
+      "reward_std": 1.4127871990203857,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.625,
+      "rewards/probe_completion_length/std": 0.49186936020851135,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.53125,
+      "rewards/probe_shaping_dominance/std": 0.507007360458374,
+      "rewards/probe_terminal_raw/mean": 0.46875,
+      "rewards/probe_terminal_raw/std": 0.507007360458374,
+      "rewards/rollout_reward_func/mean": -0.0625,
+      "rewards/rollout_reward_func/std": 1.014014720916748,
+      "sampling/importance_sampling_ratio/max": 1.0697797536849976,
+      "sampling/importance_sampling_ratio/mean": 0.8737754821777344,
+      "sampling/importance_sampling_ratio/min": 0.5936935544013977,
+      "sampling/sampling_logp_difference/max": 0.6156964302062988,
+      "sampling/sampling_logp_difference/mean": 0.07654178887605667,
+      "step": 295,
+      "step_time": 10.554522554004507
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 0.8588508367538452,
+      "epoch": 0.00296,
+      "grad_norm": 0.0030094783287495375,
+      "kl": 1.5040524452924728,
+      "learning_rate": 7.999974971828538e-06,
+      "loss": 0.0,
+      "step": 296,
+      "step_time": 6.430585491005331
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.3125,
+      "completions/mean_terminated_length": 2.3125,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 1.2149087339639664,
+      "epoch": 0.00297,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0011427103308960795,
+      "kl": 0.8978012688457966,
+      "learning_rate": 7.999974778934173e-06,
+      "loss": -0.0001,
+      "num_tokens": 7124963.0,
+      "reward": 1.7000000476837158,
+      "reward_std": 1.2443419694900513,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.3125,
+      "rewards/probe_completion_length/std": 0.4709290862083435,
+      "rewards/probe_invalid_count/mean": 0.0,
+      "rewards/probe_invalid_count/std": 0.0,
+      "rewards/probe_shaping_dominance/mean": 0.78125,
+      "rewards/probe_shaping_dominance/std": 0.420013427734375,
+      "rewards/probe_terminal_raw/mean": 0.21875,
+      "rewards/probe_terminal_raw/std": 0.420013427734375,
+      "rewards/rollout_reward_func/mean": -0.5625,
+      "rewards/rollout_reward_func/std": 0.84002685546875,
+      "sampling/importance_sampling_ratio/max": 0.9789729118347168,
+      "sampling/importance_sampling_ratio/mean": 0.7941563725471497,
+      "sampling/importance_sampling_ratio/min": 0.34077003598213196,
+      "sampling/sampling_logp_difference/max": 0.567535400390625,
+      "sampling/sampling_logp_difference/mean": 0.11595000326633453,
+      "step": 297,
+      "step_time": 10.22498737999922
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "entropy": 1.210085205733776,
+      "epoch": 0.00298,
+      "grad_norm": 0.0011137030087411404,
+      "kl": 0.897330904379487,
+      "learning_rate": 7.999974585299335e-06,
+      "loss": -0.0001,
+      "step": 298,
+      "step_time": 6.39810065599886
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.0,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.0,
+      "completions/clipped_ratio": 0.0,
+      "completions/max_length": 3.0,
+      "completions/max_terminated_length": 3.0,
+      "completions/mean_length": 2.75,
+      "completions/mean_terminated_length": 2.75,
+      "completions/min_length": 2.0,
+      "completions/min_terminated_length": 2.0,
+      "entropy": 0.9047636911273003,
+      "epoch": 0.00299,
+      "frac_reward_zero_std": 0.75,
+      "grad_norm": 0.0018326605204492807,
+      "kl": 1.868786558508873,
+      "learning_rate": 7.999974390924023e-06,
+      "loss": -0.0001,
+      "num_tokens": 7174714.0,
+      "reward": 1.9812500476837158,
+      "reward_std": 0.8607714176177979,
+      "rewards/format_guard/mean": -0.05000000074505806,
+      "rewards/format_guard/std": 0.0,
+      "rewards/probe_completion_length/mean": 1.75,
+      "rewards/probe_completion_length/std": 0.4399413466453552,
+      "rewards/probe_invalid_count/mean": 0.03125,
+      "rewards/probe_invalid_count/std": 0.1767766922712326,
+      "rewards/probe_shaping_dominance/mean": 0.875,
+      "rewards/probe_shaping_dominance/std": 0.33601075410842896,
+      "rewards/probe_terminal_raw/mean": 0.125,
+      "rewards/probe_terminal_raw/std": 0.33601075410842896,
+      "rewards/rollout_reward_func/mean": -0.75,
+      "rewards/rollout_reward_func/std": 0.6720215082168579,
+      "sampling/importance_sampling_ratio/max": 1.6604421138763428,
+      "sampling/importance_sampling_ratio/mean": 0.8135799765586853,
+      "sampling/importance_sampling_ratio/min": 0.07729801535606384,
+      "sampling/sampling_logp_difference/max": 2.0219509601593018,
+      "sampling/sampling_logp_difference/mean": 0.16112717986106873,
+      "step": 299,
+      "step_time": 10.90222747000007
+    },
+    {
+      "clip_ratio/high_max": 0.0,
+      "clip_ratio/high_mean": 0.0,
+      "clip_ratio/low_mean": 0.02083333395421505,
+      "clip_ratio/low_min": 0.0,
+      "clip_ratio/region_mean": 0.02083333395421505,
+      "entropy": 0.90055151283741,
+      "epoch": 0.003,
+      "grad_norm": 0.0011463422561064363,
+      "kl": 1.9187089204788208,
+      "learning_rate": 7.999974195808239e-06,
+      "loss": -0.0001,
+      "step": 300,
+      "step_time": 6.1042288429962355
     }
   ],
   "logging_steps": 1.0,
   "max_steps": 200000,
-  "num_input_tokens_seen": 1769976,
+  "num_input_tokens_seen": 7174714,
   "num_train_epochs": 2,
   "save_steps": 500,
   "stateful_callbacks": {
@@ -2216,7 +8727,7 @@
       "attributes": {}
     }
   },
-  "total_flos": 0.0,
+  "total_flos": 1898.0,
   "train_batch_size": 2,
   "trial_name": null,
   "trial_params": null