Views
No views yet
erdos. Checkpoint saved
after training step 47 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 47,
3 "progress/batch": 47,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.96,
6 "puct/buffer_size": 756,
7 "puct/sampled_size": 8,
8 "puct/T": 24064,
9 "puct/scale_last": 0.11905935396483913,
10 "puct/buffer_value/mean": -0.38303298666324925,
11 "puct/buffer_value/std": 0.014511519360143155,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38094064603516675,
14 "puct/buffer_timestep/mean": 22.66931216931217,
15 "puct/buffer_timestep/std": 13.703749923314836,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 46.0,
18 "puct/buffer_construction_len/mean": 79.42724867724868,
19 "puct/buffer_construction_len/std": 8.112551405927793,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.3809407829615887,
23 "puct/sampled_value/std": 7.670664694549943e-08,
24 "puct/sampled_value/min": -0.3809408282278517,
25 "puct/sampled_value/max": -0.38094064603516675,
26 "puct/sampled_timestep/mean": 46.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 46.0,
29 "puct/sampled_timestep/max": 46.0,
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": 2964.8935174942017,
35 "env/all/ac_tokens_per_turn": 9539.759765625,
36 "env/all/ob_tokens_per_turn": 1274.25,
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": 4884357,
41 "env/all/total_ob_tokens": 652416,
42 "env/all/time/sampling_mean": 299.3585044145584,
43 "env/all/time/sampling_max": 438.3633146286011,
44 "env/all/time/env_step_mean": 980.5848724837415,
45 "env/all/time/env_step_max": 2528.596714735031,
46 "env/all/reward/mean": 0.4452104681344051,
47 "env/all/reward/max": 2.6250811834991308,
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": 0.4452104681344051,
53 "env/all/correctness": 0.169921875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.3818084355019077,
57 "env/all/raw_score/min": 0.3809405895844437,
58 "env/all/raw_score/max": 0.4564072062115783,
59 "env/all/initial_raw_score": -0.3809407829615886,
60 "env/all/initial_raw_score/min": -0.3809408282278517,
61 "env/all/initial_raw_score/max": -0.38094064603516675,
62 "env/all/msg": "RuntimeError: Program execution failed: ValueError: Please select a valid mutation strategy",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n # Only consider valid shifts\n max_corr = np.max(corr[n_points - 1:2 * n_points - 1])\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n # Initialize with a better (random) construction\n n_points = 100 # Increased to allow higher resolution\n required_sum = n_points / 2.0\n h_initial = np.random.uniform(0, 1, n_points)\n h_initial = (h_initial / np.sum(h_initial)) * required_sum\n\n dx = 2.0 / n_points\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(x):\n h = np.array(x)\n sum_h = np.sum(h)\n penalty = 1e4 * (sum_h - required_sum) ** 2 # Reduced penalty coefficient\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n result = differential_evolution(\n objective,\n bounds,\n strategy='current-to-best/1', # Better exploration\n popsize=50, # Smaller popsize for efficiency\n maxiter=600, # More iterations for higher resolution\n mutation=(0.3, 0.9), # Smaller mutation range\n recombination=0.5, # More exploitation\n tol=1e-4, # Tighter convergence\n seed=seed,\n x0=h_initial\n )\n\n best_h = result.x\n best_c5 = evaluate_c5(best_h, dx)\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 299.3585044145584,
65 "env/all/time/policy/min": 146.35565876960754,
66 "env/all/time/policy/max": 438.3633146286011,
67 "env/all/time/env_step": 980.5848724837415,
68 "env/all/time/env_step/min": 0.005625247955322266,
69 "env/all/time/env_step/max": 2528.596714735031,
70 "env/all/time/reward_compute": 3.2177194952964783e-07,
71 "env/all/time/reward_compute/min": 2.1606683731079102e-07,
72 "env/all/time/reward_compute/max": 4.6938657760620117e-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": 0.04966767877340317,
77 "advantage/min": -1.0,
78 "advantage/max": 18.954360961914062,
79 "time/assemble_training_data": 7.873572587966919,
80 "time/kl_vs_base": 86.49407267570496,
81 "kl_policy_base": 0.0009427188197150826,
82 "time/train": 549.3293282985687,
83 "time/save_checkpoint": 15.573072910308838,
84 "time/total": 3625.4596490859985
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