Views
No views yet
erdos. Checkpoint saved
after training step 43 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 43,
3 "progress/batch": 43,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.88,
6 "puct/buffer_size": 692,
7 "puct/sampled_size": 8,
8 "puct/T": 22016,
9 "puct/scale_last": 0.1190592132566401,
10 "puct/buffer_value/mean": -0.38322646958144085,
11 "puct/buffer_value/std": 0.015153150843050743,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809407867433658,
14 "puct/buffer_timestep/mean": 20.65028901734104,
15 "puct/buffer_timestep/std": 12.525659786839961,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 42.0,
18 "puct/buffer_construction_len/mean": 79.3742774566474,
19 "puct/buffer_construction_len/std": 8.477449016005439,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.3809408339536986,
23 "puct/sampled_value/std": 1.827636657939072e-08,
24 "puct/sampled_value/min": -0.3809408433864274,
25 "puct/sampled_value/max": -0.3809407867433658,
26 "puct/sampled_timestep/mean": 42.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 42.0,
29 "puct/sampled_timestep/max": 42.0,
30 "puct/sampled_construction_len/mean": 80.0,
31 "puct/sampled_construction_len/std": 0.0,
32 "puct/sampled_construction_len/min": 80.0,
33 "puct/sampled_construction_len/max": 80.0,
34 "time/sampling": 3791.197613954544,
35 "env/all/ac_tokens_per_turn": 9186.6328125,
36 "env/all/ob_tokens_per_turn": 1322.0,
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": 4703556,
41 "env/all/total_ob_tokens": 676864,
42 "env/all/time/sampling_mean": 285.06941909343004,
43 "env/all/time/sampling_max": 398.1377501487732,
44 "env/all/time/env_step_mean": 1413.7217261344194,
45 "env/all/time/env_step_max": 3403.0998730659485,
46 "env/all/reward/mean": 0.6890375136952999,
47 "env/all/reward/max": 2.6250795652084373,
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": 0.6890375136952999,
53 "env/all/correctness": 0.267578125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.389941484008357,
57 "env/all/raw_score/min": 0.38094082442404065,
58 "env/all/raw_score/max": 0.5,
59 "env/all/initial_raw_score": -0.38094083395369854,
60 "env/all/initial_raw_score/min": -0.3809408433864274,
61 "env/all/initial_raw_score/max": -0.3809407867433658,
62 "env/all/msg": "Success; raw_score=0.3809417071852029",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\n\ndef evaluate_c5(h, dx):\n \"\"\"Efficient evaluation of the maximum overlap using cross-correlation.\"\"\"\n cross_corr = np.correlate(h, 1.0 - h, mode='full')\n max_integral = np.max(cross_corr) * dx\n return max_integral\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n import random\n random.seed(seed)\n\n # Use provided initial_h_values if available\n if 'initial_h_values' in globals() and len(initial_h_values) > 0:\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2\n initial_guess = initial_h_values.copy()\n else:\n n_points = 80\n dx = 2.0 / n_points\n required_sum = n_points / 2\n initial_guess = np.zeros(n_points)\n initial_guess[:n_points//2] = 1.0\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(x):\n h = np.array(x)\n max_c5 = evaluate_c5(h, dx)\n sum_h = np.sum(h)\n penalty = 100 * (sum_h - required_sum) ** 2 # Reduced penalty coefficient\n return max_c5 + penalty\n\n # Global optimization with Differential Evolution\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='rand1bin',\n popsize=40,\n maxiter=1000,\n tol=1e-8,\n mutation=(0.2, 0.5),\n recombination=0.8,\n seed=seed,\n x0=initial_guess\n )\n\n best_h_de = result_de.x\n best_c5_de = evaluate_c5(best_h_de, dx)\n\n # Local refinement with L-BFGS-B\n result_local = minimize(\n objective,\n best_h_de,\n method='L-BFGS-B',\n bounds=bounds,\n tol=1e-8\n )\n\n best_h = result_local.x\n best_c5 = evaluate_c5(best_h, dx)\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 285.06941909343004,
65 "env/all/time/policy/min": 132.7878634929657,
66 "env/all/time/policy/max": 398.1377501487732,
67 "env/all/time/env_step": 1413.7217261344194,
68 "env/all/time/env_step/min": 0.006463766098022461,
69 "env/all/time/env_step/max": 3403.0998730659485,
70 "env/all/time/reward_compute": 3.1851232051849365e-07,
71 "env/all/time/reward_compute/min": 2.1606683731079102e-07,
72 "env/all/time/reward_compute/max": 5.587935447692871e-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": 0.031055361032485962,
77 "advantage/min": -1.0,
78 "advantage/max": 5.6269450187683105,
79 "time/assemble_training_data": 7.770634412765503,
80 "time/kl_vs_base": 77.24723529815674,
81 "kl_policy_base": 0.0009622994693927467,
82 "time/train": 528.5311992168427,
83 "time/save_checkpoint": 14.34505009651184,
84 "time/total": 4420.829270124435
85}[2026-07-09T06:24:18+00:00] job=1812624 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:31:44+00:00] job=1812955 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T07:59:00+00:00] job=1813123 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:28:32+00:00] job=1813124 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T09:35:59+00:00] job=1813609 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:45:57+00:00] job=1813622 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-10T08:20:19+00:00] job=1813610 node=node-29 ngpu=6 ntrain=2 replicas=4 flash_attn=yes