Views
No views yet
erdos. Checkpoint saved
after training step 3 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 3,
3 "progress/batch": 3,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.08,
6 "puct/buffer_size": 56,
7 "puct/sampled_size": 8,
8 "puct/T": 1536,
9 "puct/scale_last": 0.030004933342652362,
10 "puct/buffer_value/mean": -0.4015208165132749,
11 "puct/buffer_value/std": 0.04080882206957126,
12 "puct/buffer_value/min": -0.5130522804051018,
13 "puct/buffer_value/max": -0.38137472041478315,
14 "puct/buffer_timestep/mean": 0.7142857142857143,
15 "puct/buffer_timestep/std": 1.0301575072754257,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 2.0,
18 "puct/buffer_construction_len/mean": 67.67857142857143,
19 "puct/buffer_construction_len/std": 23.059788085119347,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 100.0,
22 "puct/sampled_value/mean": -0.3816836391431756,
23 "puct/sampled_value/std": 0.00014957053549863883,
24 "puct/sampled_value/min": -0.3818520388530266,
25 "puct/sampled_value/max": -0.38137472041478315,
26 "puct/sampled_timestep/mean": 2.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 2.0,
29 "puct/sampled_timestep/max": 2.0,
30 "puct/sampled_construction_len/mean": 61.25,
31 "puct/sampled_construction_len/std": 22.431841208425134,
32 "puct/sampled_construction_len/min": 40.0,
33 "puct/sampled_construction_len/max": 99.0,
34 "time/sampling": 934.6010000705719,
35 "env/all/ac_tokens_per_turn": 9398.125,
36 "env/all/ob_tokens_per_turn": 1226.75,
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": 4811840,
41 "env/all/total_ob_tokens": 628096,
42 "env/all/time/sampling_mean": 515.4491372443736,
43 "env/all/time/sampling_max": 768.6091659069061,
44 "env/all/time/env_step_mean": 14.849666357040405,
45 "env/all/time/env_step_max": 238.18644762039185,
46 "env/all/reward/mean": 0.9313095851005242,
47 "env/all/reward/max": 2.622148362290573,
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": 0.9313095851005242,
53 "env/all/correctness": 0.369140625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.39916791532027185,
57 "env/all/raw_score/min": 0.38136666412916786,
58 "env/all/raw_score/max": 0.522073486654163,
59 "env/all/initial_raw_score": -0.3816836391431756,
60 "env/all/initial_raw_score/min": -0.3818520388530266,
61 "env/all/initial_raw_score/max": -0.38137472041478315,
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, minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n global h_initial\n h_initial = initial_h_values # Provided by external context\n n_points = len(h_initial)\n dx = 2.0 / n_points\n\n # Objective function to minimize the maximum overlap\n def objective(h):\n h_array = np.array(h, dtype=np.float64)\n # Adjust the sum to meet the constraint\n sum_h = np.sum(h_array)\n delta = (n_points / 2) - sum_h\n h_adj = h_array.copy()\n h_adj[-1] += delta\n h_adj[-1] = np.clip(h_adj[-1], 0.0, 1.0)\n # Compute the overlap integral\n corr = np.correlate(h_adj, 1 - h_adj, mode=\"full\")\n return np.max(corr) * dx\n\n # Run Differential Evolution for global exploration\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n result_de = differential_evolution(\n fun=objective,\n bounds=bounds,\n strategy='best1bin',\n popsize=30,\n mutation=(0.5, 1.0),\n recombination=0.8,\n tol=1e-8,\n maxiter=120,\n polish=True,\n vectorized=True\n )\n\n # Refine the best candidate using SLSQP\n def objective_slsqp(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n return np.max(corr) * dx\n\n constraints = [\n {'type': 'eq', 'fun': lambda h: np.sum(h) - n_points / 2}\n ]\n\n bounds_slsqp = [(0.0, 1.0) for _ in range(n_points)]\n res_slsqp = minimize(\n fun=objective_slsqp,\n x0=result_de.x,\n method='SLSQP',\n bounds=bounds_slsqp,\n constraints=constraints,\n options={'ftol': 1e-9, 'maxiter': 300}\n )\n\n return res_slsqp.x, res_slsqp.fun, n_points\n```",
64 "env/all/time/policy": 515.4491372443736,
65 "env/all/time/policy/min": 236.38408851623535,
66 "env/all/time/policy/max": 768.6091659069061,
67 "env/all/time/env_step": 14.849666357040405,
68 "env/all/time/env_step/min": 0.005922079086303711,
69 "env/all/time/env_step/max": 238.18644762039185,
70 "env/all/time/reward_compute": 4.7031790018081665e-07,
71 "env/all/time/reward_compute/min": 2.7567148208618164e-07,
72 "env/all/time/reward_compute/max": 9.499490261077881e-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": 0.023142989724874496,
77 "advantage/min": -0.998773992061615,
78 "advantage/max": 2.851428747177124,
79 "time/assemble_training_data": 8.55763292312622,
80 "time/kl_vs_base": 128.02790307998657,
81 "kl_policy_base": 0.0005834384937770665,
82 "time/train": 984.4880492687225,
83 "time/save_checkpoint": 7.718710422515869,
84 "time/total": 2064.4912378787994
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