Views
No views yet
erdos. Checkpoint saved
after training step 39 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 39,
3 "progress/batch": 39,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.8,
6 "puct/buffer_size": 632,
7 "puct/sampled_size": 8,
8 "puct/T": 19968,
9 "puct/scale_last": 0.279076682489534,
10 "puct/buffer_value/mean": -0.3838636046793858,
11 "puct/buffer_value/std": 0.0189751733537608,
12 "puct/buffer_value/min": -0.66,
13 "puct/buffer_value/max": -0.38092331751046604,
14 "puct/buffer_timestep/mean": 18.746835443037973,
15 "puct/buffer_timestep/std": 11.404494288901162,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 38.0,
18 "puct/buffer_construction_len/mean": 94.77215189873418,
19 "puct/buffer_construction_len/std": 31.472089047920186,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 200.0,
22 "puct/sampled_value/mean": -0.38092373428170617,
23 "puct/sampled_value/std": 2.0222440347431146e-07,
24 "puct/sampled_value/min": -0.3809240377282457,
25 "puct/sampled_value/max": -0.38092331751046604,
26 "puct/sampled_timestep/mean": 38.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 38.0,
29 "puct/sampled_timestep/max": 38.0,
30 "puct/sampled_construction_len/mean": 100.0,
31 "puct/sampled_construction_len/std": 0.0,
32 "puct/sampled_construction_len/min": 100.0,
33 "puct/sampled_construction_len/max": 100.0,
34 "time/sampling": 1881.9964051246643,
35 "env/all/ac_tokens_per_turn": 9887.65234375,
36 "env/all/ob_tokens_per_turn": 1385.875,
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": 5062478,
41 "env/all/total_ob_tokens": 709568,
42 "env/all/time/sampling_mean": 548.5941278086975,
43 "env/all/time/sampling_max": 873.0590214729309,
44 "env/all/time/env_step_mean": 75.06931490032002,
45 "env/all/time/env_step_max": 1100.1414487361908,
46 "env/all/reward/mean": 0.6431065269234666,
47 "env/all/reward/max": 2.6253937248670174,
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": 0.6431065269234666,
53 "env/all/correctness": 0.26953125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.4308486006982775,
57 "env/all/raw_score/min": 0.3808952403117042,
58 "env/all/raw_score/max": 1.0,
59 "env/all/initial_raw_score": -0.3809237342817062,
60 "env/all/initial_raw_score/min": -0.3809240377282457,
61 "env/all/initial_raw_score/max": -0.38092331751046604,
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 = 100 # Small number of samples for performance\n dx = 2.0 / n_points\n desired_sum = n_points / 2.0 # Normalization constant for area 1\n\n best_t = 0.5\n best_h = np.zeros(n_points)\n\n if 'initial_h_values' in globals():\n initial_h = np.array(initial_h_values)\n if initial_h.shape[0] > n_points:\n initial_h = initial_h[::2] # Downsample\n elif initial_h.shape[0] < n_points:\n # Upsample by repeating\n initial_h = np.repeat(initial_h, n_points // len(initial_h))\n initial_h = initial_h[:n_points]\n # Adjust sum and clip to [0, 1]\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 # Fallback to random\n initial_h = np.random.rand(n_points)\n initial_h = initial_h / np.sum(initial_h) * desired_sum\n initial_h = np.clip(initial_h, 0, 1)\n\n # Objective: minimize the upper bound t\n def objective(x):\n h_vec = x[:-1]\n t_val = x[-1]\n cross_corr = np.correlate(h_vec, 1 - h_vec, mode='full')\n overlaps = cross_corr * dx\n current_max = np.max(overlaps)\n return current_max\n\n def constraint_sum(x):\n return np.sum(x[:-1]) - desired_sum\n\n def constraints_func(x):\n h_vec = x[:-1]\n cross_corr = np.correlate(h_vec, 1 - h_vec, mode='full')\n overlaps = cross_corr * dx\n return [x[-1] - np.max(overlaps)] # Single constraint: t \u2265 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((initial_h, [0.5])),\n method='L-BFGS-B',\n bounds=bounds,\n constraints=cons,\n options={\n 'ftol': 1e-6, # Reduced tolerance for faster convergence\n 'maxiter': 200, # Reasonable iteration limit\n 'disp': False\n }\n )\n if result.success:\n current_t = result.x[-1]\n if current_t < best_t:\n best_t = current_t\n best_h = result.x[:-1]\n except:\n pass\n\n return best_h, best_t, n_points\n```",
64 "env/all/time/policy": 548.5941278086975,
65 "env/all/time/policy/min": 231.95224523544312,
66 "env/all/time/policy/max": 873.0590214729309,
67 "env/all/time/env_step": 75.06931490032002,
68 "env/all/time/env_step/min": 0.007306814193725586,
69 "env/all/time/env_step/max": 1100.1414487361908,
70 "env/all/time/reward_compute": 3.948807716369629e-07,
71 "env/all/time/reward_compute/min": 2.7194619178771973e-07,
72 "env/all/time/reward_compute/max": 8.568167686462402e-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": 0.03134578466415405,
77 "advantage/min": -0.9660258293151855,
78 "advantage/max": 7.257862091064453,
79 "time/assemble_training_data": 6.804502010345459,
80 "time/kl_vs_base": 136.9415442943573,
81 "kl_policy_base": 0.0008389363647438586,
82 "time/train": 1057.8202304840088,
83 "time/save_checkpoint": 18.67066478729248,
84 "time/total": 3103.5269503593445
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