Views
No views yet
erdos. Checkpoint saved
after training step 16 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 16,
3 "progress/batch": 16,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.34,
6 "puct/buffer_size": 264,
7 "puct/sampled_size": 8,
8 "puct/T": 8192,
9 "puct/scale_last": 0.11905844323699583,
10 "puct/buffer_value/mean": -0.3862092449775404,
11 "puct/buffer_value/std": 0.023103722062688424,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38094155676301006,
14 "puct/buffer_timestep/mean": 7.242424242424242,
15 "puct/buffer_timestep/std": 4.767505561041729,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 15.0,
18 "puct/buffer_construction_len/mean": 78.2840909090909,
19 "puct/buffer_construction_len/std": 13.59947820715512,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.3809416031787699,
23 "puct/sampled_value/std": 1.846224591702292e-08,
24 "puct/sampled_value/min": -0.3809416173591929,
25 "puct/sampled_value/max": -0.38094155676301006,
26 "puct/sampled_timestep/mean": 15.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 15.0,
29 "puct/sampled_timestep/max": 15.0,
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": 1745.9462115764618,
35 "env/all/ac_tokens_per_turn": 9363.658203125,
36 "env/all/ob_tokens_per_turn": 1702.75,
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": 4794193,
41 "env/all/total_ob_tokens": 871808,
42 "env/all/time/sampling_mean": 509.63133732927963,
43 "env/all/time/sampling_max": 747.7370183467865,
44 "env/all/time/env_step_mean": 25.222510970197618,
45 "env/all/time/env_step_max": 1100.1218929290771,
46 "env/all/reward/mean": 0.9626485443774048,
47 "env/all/reward/max": 2.625074518651347,
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": 0.9626485443774048,
53 "env/all/correctness": 0.37890625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.39654633009992707,
57 "env/all/raw_score/min": 0.3809415567612202,
58 "env/all/raw_score/max": 0.5183051091845526,
59 "env/all/initial_raw_score": -0.38094160317876985,
60 "env/all/initial_raw_score/min": -0.3809416173591929,
61 "env/all/initial_raw_score/max": -0.38094155676301006,
62 "env/all/msg": "RuntimeError: Program execution failed: TypeError: differential_evolution() got an unexpected keyword argument 'fun'",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\nfrom scipy.signal import correlate\n\n# Pre-imported functions\n# from evaluate_erdos_solution import evaluate_C5\n# from initial_h_values import initial_h_values\n\ndef generate_initial_guesses(n_points, required_sum):\n \"\"\"\n Generate a diverse set of initial guesses for the step function h.\n \"\"\"\n guesses = []\n\n # Use the current best initial_h_values\n if 'initial_h_values' in globals():\n h = initial_h_values.copy()\n h = h / np.sum(h) * required_sum\n guesses.append(h)\n\n # Perturbed version of the current solution\n if 'initial_h_values' in globals():\n h_perturbed = initial_h_values.copy() + np.random.normal(0, 0.05, n_points)\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n h_perturbed = h_perturbed / np.sum(h_perturbed) * required_sum\n guesses.append(h_perturbed)\n\n # Random binary patterns\n for _ in range(10):\n indices = np.random.choice(n_points, size=required_sum, replace=False)\n h_random = np.zeros(n_points)\n h_random[indices] = 1.0\n guesses.append(h_random)\n\n # Evenly spaced binary pattern\n h_even = np.zeros(n_points)\n for i in range(n_points):\n if i % 2 == 0:\n h_even[i] = 1.0\n h_even = h_even / np.sum(h_even) * required_sum\n guesses.append(h_even)\n\n # Alternating blocks of 1s and 0s\n h_alternate = np.zeros(n_points)\n for i in range(n_points):\n if i % 4 == 0 or i % 4 == 1:\n h_alternate[i] = 1.0\n h_alternate = h_alternate / np.sum(h_alternate) * required_sum\n guesses.append(h_alternate)\n\n # Cluster-based patterns (3, 4, and 5 clusters)\n for n_clusters in [3, 4, 5]:\n centers = np.random.choice(n_points, size=n_clusters, replace=False)\n h_cluster = np.zeros(n_points)\n for c in centers:\n h_cluster[c] = 1.0\n h_cluster = h_cluster / np.sum(h_cluster) * required_sum\n guesses.append(h_cluster)\n\n # Random non-binary pattern\n h_rand_nonbinary = np.random.rand(n_points)\n h_rand_nonbinary = h_rand_nonbinary / np.sum(h_rand_nonbinary) * required_sum\n guesses.append(h_rand_nonbinary)\n\n # Random permutation of 1s and 0s\n h_permutation = np.zeros(n_points)\n h_permutation[:required_sum] = 1.0\n np.random.shuffle(h_permutation)\n h_permutation = h_permutation / np.sum(h_permutation) * required_sum\n guesses.append(h_permutation)\n\n return guesses\n\ndef evaluate_C5(h, dx):\n h1 = 1.0 - h\n corr = correlate(h, h1, mode='full')\n max_corr = np.max(corr)\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n import random\n random.seed(seed)\n\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = int(n_points / 2)\n\n best_h = None\n best_c5 = float('inf')\n\n initial_guesses = generate_initial_guesses(n_points, required_sum)\n\n def objective_function(x):\n c5_val = evaluate_C5(x, dx)\n penalty = 1e6 * (np.sum(x) - required_sum)**2\n return c5_val + penalty\n\n for guess in initial_guesses:\n # Normalize to ensure sum constraint\n guess_sum = np.sum(guess)\n if abs(guess_sum - required_sum) > 1e-5:\n guess = guess / guess_sum * required_sum\n\n result = differential_evolution(\n fun=objective_function,\n bounds=[(0.0, 1.0) for _ in range(n_points)],\n popsize=20,\n maxiter=200,\n tolerance=1e-8,\n strategy='best1bin',\n disp=False\n )\n current_c5 = result.fun\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_h = result.x\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 509.63133732927963,
65 "env/all/time/policy/min": 252.51919984817505,
66 "env/all/time/policy/max": 747.7370183467865,
67 "env/all/time/env_step": 25.222510970197618,
68 "env/all/time/env_step/min": 0.005461692810058594,
69 "env/all/time/env_step/max": 1100.1218929290771,
70 "env/all/time/reward_compute": 3.1385570764541626e-07,
71 "env/all/time/reward_compute/min": 2.0489096641540527e-07,
72 "env/all/time/reward_compute/max": 4.1350722312927246e-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": 0.023529646918177605,
77 "advantage/min": -1.0,
78 "advantage/max": 4.881266117095947,
79 "time/assemble_training_data": 7.337669610977173,
80 "time/kl_vs_base": 127.35285520553589,
81 "kl_policy_base": 0.000824949296656996,
82 "time/train": 1037.7615365982056,
83 "time/save_checkpoint": 8.302878618240356,
84 "time/total": 2927.8875885009766
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