Views
No views yet
erdos. Checkpoint saved
after training step 14 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 14,
3 "progress/batch": 14,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.3,
6 "puct/buffer_size": 232,
7 "puct/sampled_size": 8,
8 "puct/T": 7168,
9 "puct/scale_last": 0.11905839627050108,
10 "puct/buffer_value/mean": -0.38692623835592316,
11 "puct/buffer_value/std": 0.024559017085336234,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809416037295048,
14 "puct/buffer_timestep/mean": 6.241379310344827,
15 "puct/buffer_timestep/std": 4.190754786874301,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 13.0,
18 "puct/buffer_construction_len/mean": 78.04741379310344,
19 "puct/buffer_construction_len/std": 14.491150457689544,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38094167738110274,
23 "puct/sampled_value/std": 4.447162686680136e-08,
24 "puct/sampled_value/min": -0.3809417327729232,
25 "puct/sampled_value/max": -0.3809416037295048,
26 "puct/sampled_timestep/mean": 13.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 13.0,
29 "puct/sampled_timestep/max": 13.0,
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": 1814.5249514579773,
35 "env/all/ac_tokens_per_turn": 9453.890625,
36 "env/all/ob_tokens_per_turn": 1602.375,
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": 4840392,
41 "env/all/total_ob_tokens": 820416,
42 "env/all/time/sampling_mean": 530.9087074976414,
43 "env/all/time/sampling_max": 780.2013051509857,
44 "env/all/time/env_step_mean": 39.390933700371534,
45 "env/all/time/env_step_max": 1100.145405292511,
46 "env/all/reward/mean": 0.8498177325908731,
47 "env/all/reward/max": 2.625074205689513,
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": 0.8498177325908731,
53 "env/all/correctness": 0.337890625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.4010291147487932,
57 "env/all/raw_score/min": 0.38094160217714446,
58 "env/all/raw_score/max": 0.5243023194018515,
59 "env/all/initial_raw_score": -0.38094167738110285,
60 "env/all/initial_raw_score/min": -0.3809417327729232,
61 "env/all/initial_raw_score/max": -0.3809416037295048,
62 "env/all/msg": "RuntimeError: Program execution failed: NameError: name 'Bounds' is not defined",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\nfrom scipy.signal import correlate\n\ndef generate_initial_guesses(n_points, required_sum):\n guesses = []\n\n # Use current solution as reference\n h = initial_h_values.copy()\n h /= np.sum(h)\n h *= required_sum\n guesses.append(h)\n\n # Perturbed version\n perturbed = h + np.random.normal(0, 0.03, n_points)\n perturbed = np.clip(perturbed, 0.0, 1.0)\n perturbed /= np.sum(perturbed)\n perturbed *= required_sum\n guesses.append(perturbed)\n\n # Random uniform binary pattern\n binary_indices = np.random.choice(n_points, size=required_sum, replace=False)\n h_binary = np.zeros(n_points)\n h_binary[binary_indices] = 1.0\n guesses.append(h_binary)\n\n # Even distribution of 1s\n h_even = np.zeros(n_points)\n for i in range(n_points):\n if i < required_sum:\n h_even[i] = 1.0\n h_even /= np.sum(h_even)\n 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 in [0, 1]:\n h_alternate[i] = 1.0\n h_alternate /= np.sum(h_alternate)\n h_alternate *= required_sum\n guesses.append(h_alternate)\n\n # Half-ones pattern\n h_half = np.zeros(n_points)\n h_half[:required_sum] = 1.0\n guesses.append(h_half)\n\n # Random cluster-based pattern\n clusters = np.random.choice(n_points, size=3, replace=False)\n h_cluster = np.zeros(n_points)\n h_cluster[clusters] = 1.0\n h_cluster /= np.sum(h_cluster)\n h_cluster *= required_sum\n guesses.append(h_cluster)\n\n return guesses\n\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\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) + (1 if n_points % 2 else 0)\n\n initial_guesses = generate_initial_guesses(n_points, required_sum)\n best_h = None\n best_c5 = float('inf')\n\n for guess in initial_guesses:\n result = minimize(\n fun=lambda x: evaluate_C5(x, dx),\n x0=guess,\n method=\"L-BFGS-B\",\n bounds=Bounds(0.0, 1.0),\n constraints=[{'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum}],\n options={\n \"ftol\": 1e-10,\n \"gtol\": 1e-10,\n \"maxiter\": 500,\n \"disp\": False\n }\n )\n if result.success:\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": 530.9087074976414,
65 "env/all/time/policy/min": 263.8343403339386,
66 "env/all/time/policy/max": 780.2013051509857,
67 "env/all/time/env_step": 39.390933700371534,
68 "env/all/time/env_step/min": 0.005780458450317383,
69 "env/all/time/env_step/max": 1100.145405292511,
70 "env/all/time/reward_compute": 4.87547367811203e-07,
71 "env/all/time/reward_compute/min": 2.7567148208618164e-07,
72 "env/all/time/reward_compute/max": 9.08970832824707e-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": 0.025965094566345215,
77 "advantage/min": -1.0,
78 "advantage/max": 4.252557277679443,
79 "time/assemble_training_data": 9.479153394699097,
80 "time/kl_vs_base": 136.4198169708252,
81 "kl_policy_base": 0.0008236804278567433,
82 "time/train": 1035.5127620697021,
83 "time/save_checkpoint": 13.057955503463745,
84 "time/total": 3010.081531763077
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