Views
No views yet
erdos. Checkpoint saved
after training step 37 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 37,
3 "progress/batch": 37,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.76,
6 "puct/buffer_size": 600,
7 "puct/sampled_size": 8,
8 "puct/T": 18944,
9 "puct/scale_last": 0.2790759620810632,
10 "puct/buffer_value/mean": -0.3838565744801744,
11 "puct/buffer_value/std": 0.019073553935541426,
12 "puct/buffer_value/min": -0.66,
13 "puct/buffer_value/max": -0.3809240379189368,
14 "puct/buffer_timestep/mean": 17.746666666666666,
15 "puct/buffer_timestep/std": 10.8272413640574,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 36.0,
18 "puct/buffer_construction_len/mean": 94.59333333333333,
19 "puct/buffer_construction_len/std": 32.20053760765838,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 200.0,
22 "puct/sampled_value/mean": -0.38092467429155064,
23 "puct/sampled_value/std": 5.168093392430964e-07,
24 "puct/sampled_value/min": -0.3809252479572391,
25 "puct/sampled_value/max": -0.3809240379189368,
26 "puct/sampled_timestep/mean": 36.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 36.0,
29 "puct/sampled_timestep/max": 36.0,
30 "puct/sampled_construction_len/mean": 112.5,
31 "puct/sampled_construction_len/std": 33.071891388307385,
32 "puct/sampled_construction_len/min": 100.0,
33 "puct/sampled_construction_len/max": 200.0,
34 "time/sampling": 1692.0111830234528,
35 "env/all/ac_tokens_per_turn": 9744.923828125,
36 "env/all/ob_tokens_per_turn": 1507.75,
37 "env/all/turns_per_episode": 1.0,
38 "env/all/total_episodes": 512,
39 "env/all/total_turns": 512,
40 "env/all/total_ac_tokens": 4989401,
41 "env/all/total_ob_tokens": 771968,
42 "env/all/time/sampling_mean": 561.3092758953571,
43 "env/all/time/sampling_max": 808.39430975914,
44 "env/all/time/env_step_mean": 32.97359395865351,
45 "env/all/time/env_step_max": 1100.1273522377014,
46 "env/all/reward/mean": 0.6827046152604712,
47 "env/all/reward/max": 2.625197786660391,
48 "env/all/reward/min": 0.0,
49 "env/all/format": 1.0,
50 "env/all/format/min": 1.0,
51 "env/all/format/max": 1.0,
52 "env/all/reward": 0.6827046152604712,
53 "env/all/correctness": 0.27734375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.41394783489881626,
57 "env/all/raw_score/min": 0.3809236693819395,
58 "env/all/raw_score/max": 1.0,
59 "env/all/initial_raw_score": -0.3809246742915507,
60 "env/all/initial_raw_score/min": -0.3809252479572391,
61 "env/all/initial_raw_score/max": -0.3809240379189368,
62 "env/all/msg": "Success; raw_score=0.39407944085632346",
63 "env/all/parsed_code": "```python\nimport numpy as np\nimport time\n\n\ndef compute_c5(h, n_points):\n dx = 2.0 / n_points\n cross_corr = np.correlate(h, 1 - h, mode='full')\n cross_corr = cross_corr * dx\n return np.max(cross_corr)\n\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 100\n desired_sum = n_points / 2.0\n dx = 2.0 / n_points\n\n # Initialize from the precomputed initial_h_values\n initial_h = np.array(initial_h_values, dtype=np.float64)\n initial_h = np.clip(initial_h, 0.0, 1.0)\n current_h = initial_h.copy()\n current_sum = np.sum(current_h)\n\n # Adjust the sum if needed\n if abs(current_sum - desired_sum) > 1e-6:\n delta = desired_sum - current_sum\n current_h[-1] += delta\n current_h = np.clip(current_h, 0.0, 1.0)\n\n current_c5 = compute_c5(current_h, n_points)\n best_h = current_h.copy()\n best_c5 = current_c5\n\n # Simulated Annealing Parameters\n T_initial = 1.0\n cooling_rate = 0.998\n max_iterations = 100000\n start_time = time.time()\n\n for i in range(max_iterations):\n # Perturb one random point\n idx = np.random.randint(0, n_points)\n delta = np.random.uniform(-0.1, 0.1)\n new_h = current_h.copy()\n new_h[idx] += delta\n new_h = np.clip(new_h, 0.0, 1.0)\n\n # Adjust to ensure the sum constraint is met\n new_sum = np.sum(new_h)\n required_adj = desired_sum - new_sum\n if abs(required_adj) > 1e-6:\n # Adjust a random other index\n j = np.random.randint(0, n_points)\n if j != idx:\n new_h[j] += required_adj\n else:\n j = (idx + 1) % n_points\n new_h[j] += required_adj\n new_h = np.clip(new_h, 0.0, 1.0)\n\n # Compute new C\u2085\n new_c5 = compute_c5(new_h, n_points)\n\n # Update best solution probabilistically\n if new_c5 < best_c5:\n best_c5 = new_c5\n best_h = new_h.copy()\n elif np.random.random() < np.exp(-(new_c5 - best_c5) / T_initial):\n best_c5 = new_c5\n best_h = new_h.copy()\n\n # Cool down\n T_initial *= cooling_rate\n\n # Check for time budget\n if time.time() - start_time > budget_s:\n break\n\n # Update current_h to reflect recent best\n current_h = best_h.copy()\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 561.3092758953571,
65 "env/all/time/policy/min": 214.9264039993286,
66 "env/all/time/policy/max": 808.39430975914,
67 "env/all/time/env_step": 32.97359395865351,
68 "env/all/time/env_step/min": 0.006078481674194336,
69 "env/all/time/env_step/max": 1100.1273522377014,
70 "env/all/time/reward_compute": 3.2922253012657166e-07,
71 "env/all/time/reward_compute/min": 1.5273690223693848e-07,
72 "env/all/time/reward_compute/max": 5.327165126800537e-07,
73 "env/all/by_group/frac_mixed": 1.0,
74 "env/all/by_group/frac_all_good": 0.0,
75 "env/all/by_group/frac_all_bad": 0.0,
76 "advantage/mean": 0.02966400608420372,
77 "advantage/min": -0.9795393347740173,
78 "advantage/max": 5.034461975097656,
79 "time/assemble_training_data": 6.644336223602295,
80 "time/kl_vs_base": 137.06008982658386,
81 "kl_policy_base": 0.0009024626342579722,
82 "time/train": 1058.6564359664917,
83 "time/save_checkpoint": 18.058328866958618,
84 "time/total": 2913.9379856586456
85}[2026-07-09T06:24:18+00:00] job=1812626 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:02:36+00:00] job=1813125 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:32:13+00:00] job=1813126 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T10:04:31+00:00] job=1813623 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=yes