Views
No views yet
erdos. Checkpoint saved
after training step 33 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 33,
3 "progress/batch": 33,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.68,
6 "puct/buffer_size": 535,
7 "puct/sampled_size": 8,
8 "puct/T": 16896,
9 "puct/scale_last": 0.1190591760561926,
10 "puct/buffer_value/mean": -0.38359984479478054,
11 "puct/buffer_value/std": 0.01646394742454639,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809408239438133,
14 "puct/buffer_timestep/mean": 15.719626168224298,
15 "puct/buffer_timestep/std": 9.66213767926886,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 32.0,
18 "puct/buffer_construction_len/mean": 79.15327102803738,
19 "puct/buffer_construction_len/std": 9.591607680597992,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38094150310555086,
23 "puct/sampled_value/std": 2.026354637591029e-08,
24 "puct/sampled_value/min": -0.38094151658754866,
25 "puct/sampled_value/max": -0.38094145132638335,
26 "puct/sampled_timestep/mean": 32.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 32.0,
29 "puct/sampled_timestep/max": 32.0,
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": 4310.711181879044,
35 "env/all/ac_tokens_per_turn": 9682.43359375,
36 "env/all/ob_tokens_per_turn": 1257.125,
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": 4957406,
41 "env/all/total_ob_tokens": 643648,
42 "env/all/time/sampling_mean": 304.6641508336179,
43 "env/all/time/sampling_max": 410.03794169425964,
44 "env/all/time/env_step_mean": 1807.865667087026,
45 "env/all/time/env_step_max": 3897.4875423908234,
46 "env/all/reward/mean": 0.6014768069690014,
47 "env/all/reward/max": 2.6250749188529308,
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": 0.6014768069690014,
53 "env/all/correctness": 0.23046875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.3835988997870568,
57 "env/all/raw_score/min": 0.380941498685385,
58 "env/all/raw_score/max": 0.499006449838208,
59 "env/all/initial_raw_score": -0.38094150310555086,
60 "env/all/initial_raw_score/min": -0.38094151658754866,
61 "env/all/initial_raw_score/max": -0.38094145132638335,
62 "env/all/msg": "RuntimeError: Program execution failed: TypeError: differential_evolution() got an unexpected keyword argument 'polish_iters'",
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 h = np.array(h)\n one_minus_h = 1.0 - h\n n_points = len(h)\n corr = correlate(h, one_minus_h, mode='full')\n positive_shifts_indices = np.arange(1, n_points)\n max_overlap = np.max(corr[positive_shifts_indices]) * dx\n return max_overlap\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n try:\n n_points = len(initial_h_values)\n except NameError:\n # Default to n_points = 100 for improved resolution\n n_points = 100\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n\n # Generate a better initial guess\n try:\n initial_guess = np.array(initial_h_values)\n except NameError:\n # Binary pattern with equal 1s and 0s to satisfy integral constraint\n indices = np.random.choice(n_points, size=n_points//2, replace=False)\n initial_guess = np.zeros(n_points)\n initial_guess[indices] = 1.0\n initial_guess = initial_guess\n\n # Scale to ensure the sum constraint is satisfied\n initial_guess = initial_guess * required_sum / np.sum(initial_guess)\n\n def objective(x):\n h = np.array(x)\n max_c5 = evaluate_c5(h, dx)\n # Use a slightly weaker penalty for constraint violation to allow exploration\n penalty = 1e5 * (np.sum(x) - required_sum) ** 2\n return max_c5 + penalty\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Run differential evolution with improved parameters\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=150, # Larger population for better coverage\n maxiter=800, # More iterations for better convergence\n tol=1e-6,\n mutation=(0.8, 1.0), # Adjusted mutation range for better exploration\n recombination=0.9,\n seed=seed,\n x0=initial_guess,\n polish=True,\n polish_iters=100\n )\n\n best_de = result_de.x\n best_c5 = evaluate_c5(best_de, dx)\n\n # Refine with L-BFGS-B method for local convergence\n with warnings.catch_warnings():\n warnings.simplefilter(\"ignore\")\n result_local = minimize(\n objective,\n best_de,\n method='L-BFGS-B',\n bounds=bounds,\n tol=1e-6\n )\n\n best_local = result_local.x\n final_c5 = evaluate_c5(best_local, dx)\n\n return best_local, final_c5, n_points\n```",
64 "env/all/time/policy": 304.6641508336179,
65 "env/all/time/policy/min": 138.88423991203308,
66 "env/all/time/policy/max": 410.03794169425964,
67 "env/all/time/env_step": 1807.865667087026,
68 "env/all/time/env_step/min": 0.007125377655029297,
69 "env/all/time/env_step/max": 3897.4875423908234,
70 "env/all/time/reward_compute": 3.2456591725349426e-07,
71 "env/all/time/reward_compute/min": 2.2351741790771484e-07,
72 "env/all/time/reward_compute/max": 4.954636096954346e-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": 0.033475469797849655,
77 "advantage/min": -1.0,
78 "advantage/max": 10.041560173034668,
79 "time/assemble_training_data": 7.644739151000977,
80 "time/kl_vs_base": 84.51267123222351,
81 "kl_policy_base": 0.000855912861879915,
82 "time/train": 556.5102462768555,
83 "time/save_checkpoint": 16.765967845916748,
84 "time/total": 4978.035395860672
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