Views
No views yet
erdos. Checkpoint saved
after training step 30 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 30,
3 "progress/batch": 30,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.62,
6 "puct/buffer_size": 488,
7 "puct/sampled_size": 8,
8 "puct/T": 15360,
9 "puct/scale_last": 0.11907473987989126,
10 "puct/buffer_value/mean": -0.3839317685549489,
11 "puct/buffer_value/std": 0.0170037643291453,
12 "puct/buffer_value/min": -0.5130522804051018,
13 "puct/buffer_value/max": -0.38092526012010874,
14 "puct/buffer_timestep/mean": 14.245901639344263,
15 "puct/buffer_timestep/std": 8.806956523308148,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 29.0,
18 "puct/buffer_construction_len/mean": 93.35245901639344,
19 "puct/buffer_construction_len/std": 35.15479169614208,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 200.0,
22 "puct/sampled_value/mean": -0.38092526660034676,
23 "puct/sampled_value/std": 7.738600620693116e-10,
24 "puct/sampled_value/min": -0.3809252675383982,
25 "puct/sampled_value/max": -0.3809252654757097,
26 "puct/sampled_timestep/mean": 29.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 29.0,
29 "puct/sampled_timestep/max": 29.0,
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": 1856.8368122577667,
35 "env/all/ac_tokens_per_turn": 9300.439453125,
36 "env/all/ob_tokens_per_turn": 1898.5,
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": 4761825,
41 "env/all/total_ob_tokens": 972032,
42 "env/all/time/sampling_mean": 538.8707484742627,
43 "env/all/time/sampling_max": 797.6498239040375,
44 "env/all/time/env_step_mean": 30.759665318299085,
45 "env/all/time/env_step_max": 1100.1201527118683,
46 "env/all/reward/mean": 0.6687395614650082,
47 "env/all/reward/max": 2.6251870292858936,
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": 0.6687395614650082,
53 "env/all/correctness": 0.26171875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.39486939197874393,
57 "env/all/raw_score/min": 0.38092523031402864,
58 "env/all/raw_score/max": 0.66,
59 "env/all/initial_raw_score": -0.38092526660034676,
60 "env/all/initial_raw_score/min": -0.3809252675383982,
61 "env/all/initial_raw_score/max": -0.3809252654757097,
62 "env/all/msg": "C5 mismatch: reported 0.53125496, computed 0.52629449",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\nimport warnings\n\n# Suppress warnings for cleaner output\nwarnings.filterwarnings(\"ignore\", category=UserWarning)\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 100\n dx = 2.0 / n_points\n desired_sum = n_points / 2.0\n\n # Initial guess (copy initial_h_values)\n initial_h = initial_h_values.copy()\n initial_h = np.clip(initial_h, 0.0, 1.0)\n \n # Adjust initial sum to meet the constraint\n delta = desired_sum - np.sum(initial_h)\n if abs(delta) > 1e-6:\n initial_h[-1] += delta\n initial_h = np.clip(initial_h, 0.0, 1.0)\n initial_h = initial_h.astype(np.float64)\n\n def objective(h):\n penalty = 1e6 * abs(np.sum(h) - desired_sum)\n corr = np.correlate(h, 1.0 - h, mode='full')\n c5 = np.max(corr) * dx\n return c5 + penalty\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Run DE optimization\n try:\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=20,\n mutation=(0.5, 1.0),\n recombination=0.9,\n maxiter=200,\n tol=1e-6,\n disp=False\n )\n best_h = result.x\n best_c5 = objective(best_h)\n\n except:\n # Fallback to initial guess\n best_h = initial_h\n best_c5 = objective(best_h)\n\n # Local optimization with SLSQP\n def objective_local(h):\n corr = np.correlate(h, 1.0 - h, mode='full')\n return np.max(corr) * dx\n\n def constraint(h):\n return np.sum(h) - desired_sum\n\n bounds_local = [(0.0, 1.0) for _ in range(n_points)]\n\n # Start SQP from the best_h found in DE\n initial_guess = best_h.copy()\n if abs(np.sum(initial_guess) - desired_sum) > 1e-6:\n delta = desired_sum - np.sum(initial_guess)\n initial_guess[-1] += delta\n initial_guess = np.clip(initial_guess, 0.0, 1.0)\n\n res = minimize(\n fun=objective_local,\n x0=initial_guess,\n method='SLSQP',\n bounds=bounds_local,\n constraints=[{'type': 'eq', 'fun': constraint}],\n options={\n 'ftol': 1e-9,\n 'maxiter': 200,\n 'disp': False\n }\n )\n\n if res.success:\n new_h = res.x\n new_c5 = objective_local(new_h)\n if new_c5 < best_c5:\n best_h = new_h\n best_c5 = new_c5\n\n return (best_h, best_c5, n_points)\n```",
64 "env/all/time/policy": 538.8707484742627,
65 "env/all/time/policy/min": 234.503502368927,
66 "env/all/time/policy/max": 797.6498239040375,
67 "env/all/time/env_step": 30.759665318299085,
68 "env/all/time/env_step/min": 0.0066645145416259766,
69 "env/all/time/env_step/max": 1100.1201527118683,
70 "env/all/time/reward_compute": 5.224719643592834e-07,
71 "env/all/time/reward_compute/min": 2.2724270820617676e-07,
72 "env/all/time/reward_compute/max": 1.4193356037139893e-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": 0.03831367567181587,
77 "advantage/min": -0.9783849716186523,
78 "advantage/max": 24.90949821472168,
79 "time/assemble_training_data": 7.204564094543457,
80 "time/kl_vs_base": 138.21639585494995,
81 "kl_policy_base": 0.0008608591742813587,
82 "time/train": 1050.1514692306519,
83 "time/save_checkpoint": 18.923516750335693,
84 "time/total": 3074.414120912552
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