Views
No views yet
erdos. Checkpoint saved
after training step 43 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 43,
3 "progress/batch": 43,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.88,
6 "puct/buffer_size": 696,
7 "puct/sampled_size": 8,
8 "puct/T": 22016,
9 "puct/scale_last": 0.279113450472389,
10 "puct/buffer_value/mean": -0.38404115761413665,
11 "puct/buffer_value/std": 0.018981385087687448,
12 "puct/buffer_value/min": -0.66,
13 "puct/buffer_value/max": -0.38088654952761103,
14 "puct/buffer_timestep/mean": 20.74712643678161,
15 "puct/buffer_timestep/std": 12.559026969049038,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 42.0,
18 "puct/buffer_construction_len/mean": 100.56896551724138,
19 "puct/buffer_construction_len/std": 38.43263044390653,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 200.0,
22 "puct/sampled_value/mean": -0.3808914197037369,
23 "puct/sampled_value/std": 3.7337854749534675e-06,
24 "puct/sampled_value/min": -0.3808951754050975,
25 "puct/sampled_value/max": -0.38088654952761103,
26 "puct/sampled_timestep/mean": 42.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 42.0,
29 "puct/sampled_timestep/max": 42.0,
30 "puct/sampled_construction_len/mean": 200.0,
31 "puct/sampled_construction_len/std": 0.0,
32 "puct/sampled_construction_len/min": 200.0,
33 "puct/sampled_construction_len/max": 200.0,
34 "time/sampling": 2021.9943053722382,
35 "env/all/ac_tokens_per_turn": 10368.248046875,
36 "env/all/ob_tokens_per_turn": 1451.875,
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": 5308543,
41 "env/all/total_ob_tokens": 743360,
42 "env/all/time/sampling_mean": 613.0022428254597,
43 "env/all/time/sampling_max": 904.2779071331024,
44 "env/all/time/env_step_mean": 163.52770265378058,
45 "env/all/time/env_step_max": 1128.0283861160278,
46 "env/all/reward/mean": 0.6589185234993362,
47 "env/all/reward/max": 2.625453649137271,
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": 0.6589185234993362,
53 "env/all/correctness": 0.267578125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.4128383647146856,
57 "env/all/raw_score/min": 0.38088654662559573,
58 "env/all/raw_score/max": 0.7434068439869645,
59 "env/all/initial_raw_score": -0.38089141970373686,
60 "env/all/initial_raw_score/min": -0.3808951754050975,
61 "env/all/initial_raw_score/max": -0.38088654952761103,
62 "env/all/msg": "Computed C5 is not finite: nan",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n n_points = 200\n dx = 2.0 / n_points\n desired_sum = n_points / 2.0\n\n best_h = np.zeros(n_points)\n best_t = 0.5\n\n # Use provided initial guess if available\n if 'initial_h_values' in globals():\n initial_h = np.array(initial_h_values)\n # Adjust to fit n_points\n if initial_h.shape[0] > n_points:\n initial_h = initial_h[::2]\n elif initial_h.shape[0] < n_points:\n initial_h = np.repeat(initial_h, n_points // len(initial_h))[:n_points]\n sum_h = np.sum(initial_h)\n delta = desired_sum - sum_h\n initial_h[-1] += delta\n initial_h = np.clip(initial_h, 0, 1)\n else:\n # Basic structured pattern\n initial_h = np.zeros(n_points)\n for i in range(n_points):\n if i % 4 == 0:\n initial_h[i] = 1.0\n sum_h = np.sum(initial_h)\n delta = desired_sum - sum_h\n initial_h[-1] += delta\n initial_h = np.clip(initial_h, 0, 1)\n\n # Generate diverse initial guesses\n initial_guesses = [initial_h]\n for _ in range(2): # Add variety by modifying the initial guess\n # Random noise perturbation\n perturb = initial_h + np.random.normal(0, 0.05, n_points)\n perturb = np.clip(perturb, 0, 1)\n sum_perturb = np.sum(perturb)\n delta = desired_sum - sum_perturb\n perturb[-1] += delta\n initial_guesses.append(perturb)\n\n # Random permutation\n perm = np.random.permutation(n_points)\n perm_h = initial_h[perm]\n sum_perm = np.sum(perm_h)\n delta = desired_sum - sum_perm\n perm_h[-1] += delta\n perm_h = np.clip(perm_h, 0, 1)\n initial_guesses.append(perm_h)\n\n # Shifted pattern\n shifted_h = np.roll(initial_h, 1)\n sum_shift = np.sum(shifted_h)\n delta = desired_sum - sum_shift\n shifted_h[-1] += delta\n shifted_h = np.clip(shifted_h, 0, 1)\n initial_guesses.append(shifted_h)\n\n # Increased density of 1s\n increased_h = initial_h.copy()\n increase_mask = increased_h == 0\n increased_h[increase_mask] = np.random.choice([0, 1], p=[0.5, 0.5], size=np.sum(increase_mask))\n sum_increased = np.sum(increased_h)\n delta = desired_sum - sum_increased\n increased_h[-1] += delta\n increased_h = np.clip(increased_h, 0, 1)\n initial_guesses.append(increased_h)\n\n # Optimize each initial guess\n for h in initial_guesses:\n def objective(x):\n return x[-1] # Minimize t\n \n def constraint_sum(x):\n return np.sum(x[:-1]) - desired_sum # Ensure integral constraint\n\n def constraints_func(x):\n h_vec = x[:-1]\n t = x[-1]\n cross_corr = np.correlate(h_vec, 1 - h_vec, mode='full')\n overlaps = cross_corr * dx\n max_overlap = np.max(overlaps)\n return [t - max_overlap]\n\n cons = [\n {'type': 'eq', 'fun': constraint_sum},\n {'type': 'ineq', 'fun': constraints_func}\n ]\n\n bounds = [(0.0, 1.0) for _ in range(n_points)] + [(0.0, np.inf)]\n \n try:\n result = minimize(\n fun=objective,\n x0=np.concatenate((h, [0.5])),\n bounds=bounds,\n constraints=cons,\n method='SLSQP',\n options={\n 'ftol': 1e-8,\n 'maxiter': 100,\n 'disp': False\n }\n )\n if result.success and result.x[-1] < best_t:\n best_t = result.x[-1]\n best_h = result.x[:-1]\n except:\n pass # Skip on failure\n\n return best_h, best_t, n_points\n```",
64 "env/all/time/policy": 613.0022428254597,
65 "env/all/time/policy/min": 267.4992139339447,
66 "env/all/time/policy/max": 904.2779071331024,
67 "env/all/time/env_step": 163.52770265378058,
68 "env/all/time/env_step/min": 0.007021427154541016,
69 "env/all/time/env_step/max": 1128.0283861160278,
70 "env/all/time/reward_compute": 3.4086406230926514e-07,
71 "env/all/time/reward_compute/min": 1.9744038581848145e-07,
72 "env/all/time/reward_compute/max": 6.146728992462158e-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": 0.031972967088222504,
77 "advantage/min": -1.0,
78 "advantage/max": 6.34797477722168,
79 "time/assemble_training_data": 8.66273546218872,
80 "time/kl_vs_base": 142.13786101341248,
81 "kl_policy_base": 0.0008245133794844151,
82 "time/train": 1120.9914379119873,
83 "time/save_checkpoint": 16.511499404907227,
84 "time/total": 3311.7374064922333
85}[2026-07-09T06:24:18+00:00] job=1812626 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:02:36+00:00] job=1813125 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:32:13+00:00] job=1813126 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T10:04:31+00:00] job=1813623 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=yes