Views
No views yet
erdos. Checkpoint saved
after training step 2 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 2,
3 "progress/batch": 2,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.06,
6 "puct/buffer_size": 40,
7 "puct/sampled_size": 8,
8 "puct/T": 1024,
9 "puct/scale_last": 0.02313354533540951,
10 "puct/buffer_value/mean": -0.40979699362830313,
11 "puct/buffer_value/std": 0.04899037041504793,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3819974267881357,
14 "puct/buffer_timestep/mean": 0.2,
15 "puct/buffer_timestep/std": 0.7483314773547883,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 1.0,
18 "puct/buffer_construction_len/mean": 76.375,
19 "puct/buffer_construction_len/std": 25.81054774699677,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38214781728550096,
23 "puct/sampled_value/std": 0.00012583914263974218,
24 "puct/sampled_value/min": -0.3823690414981055,
25 "puct/sampled_value/max": -0.3819974267881357,
26 "puct/sampled_timestep/mean": 1.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 1.0,
29 "puct/sampled_timestep/max": 1.0,
30 "puct/sampled_construction_len/mean": 73.125,
31 "puct/sampled_construction_len/std": 30.95737997634813,
32 "puct/sampled_construction_len/min": 42.0,
33 "puct/sampled_construction_len/max": 143.0,
34 "time/sampling": 1772.6573071479797,
35 "env/all/ac_tokens_per_turn": 9138.02734375,
36 "env/all/ob_tokens_per_turn": 1171.75,
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": 4678670,
41 "env/all/total_ob_tokens": 599936,
42 "env/all/time/sampling_mean": 488.0048552895896,
43 "env/all/time/sampling_max": 708.5899863243103,
44 "env/all/time/env_step_mean": 34.08752423059195,
45 "env/all/time/env_step_max": 1100.1202006340027,
46 "env/all/reward/mean": 0.799580868060573,
47 "env/all/reward/max": 2.621509866949141,
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": 0.799580868060573,
53 "env/all/correctness": 0.314453125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.39603583172201084,
57 "env/all/raw_score/min": 0.38145954985425273,
58 "env/all/raw_score/max": 0.5704427133112526,
59 "env/all/initial_raw_score": -0.38214781728550096,
60 "env/all/initial_raw_score/min": -0.3823690414981055,
61 "env/all/initial_raw_score/max": -0.3819974267881357,
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 initial_h_values\n np.random.seed(seed)\n h_initial = initial_h_values.copy()\n n_points = len(h_initial)\n dx = 2.0 / n_points\n\n # Objective function with constraint penalty\n def objective(h_values):\n h = np.array(h_values)\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n max_corr = np.max(corr)\n penalty = 500000 * abs(np.sum(h) - n_points / 2)\n return max_corr * dx + penalty\n\n # Bounds for the problem\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Step 1: Run Differential Evolution for global search\n result_de = differential_evolution(\n fun=objective,\n bounds=bounds,\n strategy='best1bin', # Effective for constrained problems\n maxiter=100, # Number of generations\n popsize=15, # Control population size\n mutation=(0.5, 1), # Mutation settings\n recombination=0.7, # Recombination probability\n tol=1e-6, # Stopping tolerance\n disp=False # Skip print statements for performance\n )\n \n best_h = result_de.x\n best_c5 = objective(best_h)\n\n # Step 2: Refine the solution with SLSQP\n def objective_slsqp(h_values):\n h = np.array(h_values)\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n return np.max(corr) * dx\n\n def constraint_sum(h_values):\n return np.sum(h_values) - n_points / 2\n\n # Perform local optimization\n result_slsqp = minimize(\n fun=objective_slsqp,\n x0=best_h, # Start from DE\u2019s best solution\n method='SLSQP',\n constraints=[{'type': 'eq', 'fun': constraint_sum}],\n bounds=bounds,\n options={'maxiter': 300, 'ftol': 1e-9, 'eps': 1e-9}\n )\n\n if result_slsqp.success and result_slsqp.fun < best_c5:\n best_c5 = result_slsqp.fun\n best_h = result_slsqp.x\n\n return (best_h.tolist(), best_c5, n_points)\n```",
64 "env/all/time/policy": 488.0048552895896,
65 "env/all/time/policy/min": 223.91157603263855,
66 "env/all/time/policy/max": 708.5899863243103,
67 "env/all/time/env_step": 34.08752423059195,
68 "env/all/time/env_step/min": 0.0053253173828125,
69 "env/all/time/env_step/max": 1100.1202006340027,
70 "env/all/time/reward_compute": 3.6600977182388306e-07,
71 "env/all/time/reward_compute/min": 1.862645149230957e-07,
72 "env/all/time/reward_compute/max": 1.1138617992401123e-06,
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": 0.02816164121031761,
77 "advantage/min": -1.0,
78 "advantage/max": 7.19879150390625,
79 "time/assemble_training_data": 8.29187822341919,
80 "time/kl_vs_base": 129.30560898780823,
81 "kl_policy_base": 0.0005741093773394823,
82 "time/train": 954.9608917236328,
83 "time/save_checkpoint": 8.016062498092651,
84 "time/total": 2874.6793570518494
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