Views
No views yet
erdos. Checkpoint saved
after training step 28 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 28,
3 "progress/batch": 28,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.58,
6 "puct/buffer_size": 456,
7 "puct/sampled_size": 8,
8 "puct/T": 14336,
9 "puct/scale_last": 0.11905872740134948,
10 "puct/buffer_value/mean": -0.3840544688785584,
11 "puct/buffer_value/std": 0.017793435259634643,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809412725986564,
14 "puct/buffer_timestep/mean": 13.24561403508772,
15 "puct/buffer_timestep/std": 8.22977201567446,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 27.0,
18 "puct/buffer_construction_len/mean": 79.00657894736842,
19 "puct/buffer_construction_len/std": 10.382274480291791,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.3809414998327546,
23 "puct/sampled_value/std": 8.611641009921368e-08,
24 "puct/sampled_value/min": -0.3809415382205823,
25 "puct/sampled_value/max": -0.3809412725986564,
26 "puct/sampled_timestep/mean": 27.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 27.0,
29 "puct/sampled_timestep/max": 27.0,
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": 3705.940351009369,
35 "env/all/ac_tokens_per_turn": 9208.306640625,
36 "env/all/ob_tokens_per_turn": 1320.625,
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": 4714653,
41 "env/all/total_ob_tokens": 676160,
42 "env/all/time/sampling_mean": 285.2602476035245,
43 "env/all/time/sampling_max": 456.81366658210754,
44 "env/all/time/env_step_mean": 1407.1476679858752,
45 "env/all/time/env_step_max": 3311.4759974479675,
46 "env/all/reward/mean": 0.2815332764048716,
47 "env/all/reward/max": 2.6250756228906855,
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": 0.2815332764048716,
53 "env/all/correctness": 0.109375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.39223529424912595,
57 "env/all/raw_score/min": 0.3809413965179647,
58 "env/all/raw_score/max": 0.7466601303276769,
59 "env/all/initial_raw_score": -0.3809414998327546,
60 "env/all/initial_raw_score/min": -0.3809415382205823,
61 "env/all/initial_raw_score/max": -0.3809412725986564,
62 "env/all/msg": "RuntimeError: Program execution failed: ValueError: initial_h_values must be provided",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\nfrom scipy.signal import correlate\n\ndef evaluate_c5(h, dx):\n one_minus_h = 1.0 - h\n corr = correlate(h, one_minus_h, mode='full')\n max_c5 = np.max(corr) * dx\n return max_c5\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n import random\n random.seed(seed)\n\n # Ensure initial_h_values is defined\n if 'initial_h_values' not in locals() or initial_h_values is None:\n raise ValueError(\"initial_h_values must be provided\")\n\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n\n # Prepare initial guess\n if initial_h_values is None:\n # Fallback: generate a random initial guess\n initial_guess = np.random.uniform(0.0, 1.0, n_points)\n else:\n initial_guess = np.array(initial_h_values)\n initial_guess = initial_guess / np.sum(initial_guess) * required_sum\n\n # Define bounds\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 return max_c5\n\n # Run differential evolution with enhanced parameters\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='rand1bin',\n popsize=200,\n maxiter=300, # More iterations\n tol=1e-6,\n mutation=(0.5, 0.9),\n recombination=0.9,\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 using L-BFGS-B method\n result_local = minimize(\n objective,\n best_h_de,\n bounds=bounds,\n method='L-BFGS-B',\n tol=1e-6\n )\n\n best_h_final = result_local.x\n best_c5_final = evaluate_c5(best_h_final, dx)\n\n return best_h_final, best_c5_final, n_points\n```",
64 "env/all/time/policy": 285.2602476035245,
65 "env/all/time/policy/min": 113.13975882530212,
66 "env/all/time/policy/max": 456.81366658210754,
67 "env/all/time/env_step": 1407.1476679858752,
68 "env/all/time/env_step/min": 0.00507807731628418,
69 "env/all/time/env_step/max": 3311.4759974479675,
70 "env/all/time/reward_compute": 3.1944364309310913e-07,
71 "env/all/time/reward_compute/min": 2.4586915969848633e-07,
72 "env/all/time/reward_compute/max": 4.805624485015869e-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": 0.046120353043079376,
77 "advantage/min": -0.6698862910270691,
78 "advantage/max": 9.887121200561523,
79 "time/assemble_training_data": 5.381299018859863,
80 "time/kl_vs_base": 80.95866513252258,
81 "kl_policy_base": 0.0009031476220116019,
82 "time/train": 528.9241044521332,
83 "time/save_checkpoint": 10.549998998641968,
84 "time/total": 4333.015373706818
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