Views
No views yet
erdos. Checkpoint saved
after training step 27 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 27,
3 "progress/batch": 27,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.56,
6 "puct/buffer_size": 440,
7 "puct/sampled_size": 8,
8 "puct/T": 13824,
9 "puct/scale_last": 0.11907473987989126,
10 "puct/buffer_value/mean": -0.3842530920835038,
11 "puct/buffer_value/std": 0.017877643079815358,
12 "puct/buffer_value/min": -0.5130522804051018,
13 "puct/buffer_value/max": -0.38092526012010874,
14 "puct/buffer_timestep/mean": 12.745454545454546,
15 "puct/buffer_timestep/std": 7.941188786248755,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 26.0,
18 "puct/buffer_construction_len/mean": 88.99090909090908,
19 "puct/buffer_construction_len/std": 30.575389822697556,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 200.0,
22 "puct/sampled_value/mean": -0.3809252713514705,
23 "puct/sampled_value/std": 3.4226855881271348e-09,
24 "puct/sampled_value/min": -0.38092527615321936,
25 "puct/sampled_value/max": -0.38092526566220564,
26 "puct/sampled_timestep/mean": 26.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 26.0,
29 "puct/sampled_timestep/max": 26.0,
30 "puct/sampled_construction_len/mean": 137.5,
31 "puct/sampled_construction_len/std": 48.41229182759271,
32 "puct/sampled_construction_len/min": 100.0,
33 "puct/sampled_construction_len/max": 200.0,
34 "time/sampling": 1865.0647180080414,
35 "env/all/ac_tokens_per_turn": 9312.3125,
36 "env/all/ob_tokens_per_turn": 1570.375,
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": 4767904,
41 "env/all/total_ob_tokens": 804032,
42 "env/all/time/sampling_mean": 527.3870373102836,
43 "env/all/time/sampling_max": 755.4097290039062,
44 "env/all/time/env_step_mean": 65.29666293505579,
45 "env/all/time/env_step_max": 1100.149849653244,
46 "env/all/reward/mean": 1.151468942258844,
47 "env/all/reward/max": 2.625186771763899,
48 "env/all/reward/min": 0.0,
49 "env/all/format": 0.998046875,
50 "env/all/format/min": 0.0,
51 "env/all/format/max": 1.0,
52 "env/all/reward": 1.151468942258844,
53 "env/all/correctness": 0.4609375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.40440311674495333,
57 "env/all/raw_score/min": 0.38092526768151386,
58 "env/all/raw_score/max": 0.5317004730807383,
59 "env/all/initial_raw_score": -0.3809252713514704,
60 "env/all/initial_raw_score/min": -0.38092527615321936,
61 "env/all/initial_raw_score/max": -0.38092526566220564,
62 "env/all/msg": "Success; raw_score=0.3809252717926112",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import basinhopping, minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 200\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Generate initial guess or fallback to structured initial\n try:\n if len(initial_h_values) == n_points:\n initial_h = initial_h_values.copy()\n else:\n m_intervals = 10\n h = np.zeros(n_points)\n L = 1.0 / m_intervals\n start = 0.0\n for i in range(m_intervals):\n start_idx = int(start / dx)\n end_idx = int((start + L) / dx)\n if start_idx < n_points and end_idx <= n_points:\n h[start_idx:end_idx] = 1.0\n start += L\n current_sum = np.sum(h)\n delta = (n_points / 2) - current_sum\n idx = np.where(h == 0)[0]\n if len(idx) > 0:\n h[idx[0]] += delta\n initial_h = h\n except NameError:\n m_intervals = 10\n h = np.zeros(n_points)\n L = 1.0 / m_intervals\n start = 0.0\n for i in range(m_intervals):\n start_idx = int(start / dx)\n end_idx = int((start + L) / dx)\n if start_idx < n_points and end_idx <= n_points:\n h[start_idx:end_idx] = 1.0\n start += L\n current_sum = np.sum(h)\n delta = (n_points / 2) - current_sum\n idx = np.where(h == 0)[0]\n if len(idx) > 0:\n h[idx[0]] += delta\n initial_h = h\n\n # Define objective function with penalty for constraint\n def objective_with_penalty(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n obj = np.max(corr) * dx\n penalty = 1e3 * (np.sum(h) - n_points / 2) ** 2\n return obj + penalty\n\n # Basinhopping with increased parameters\n result_bh = basinhopping(\n objective_with_penalty,\n initial_h,\n niter=500, # More iterative search\n T=0.5, # Lower temperature to focus\n stepsize=0.8, # Larger step for global search\n minimizer_kwargs={\n 'method': 'L-BFGS-B',\n 'bounds': [(0.0, 1.0) for _ in range(n_points)]\n },\n disp=False\n )\n\n best_h_bh = result_bh.x\n best_c5_bh = objective_with_penalty(best_h_bh)\n\n # Local SLSQP refinement\n def objective(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 result_slsqp = minimize(\n fun=objective,\n x0=best_h_bh,\n method='SLSQP',\n bounds=[(0.0, 1.0) for _ in range(n_points)],\n constraints=[dict(type='eq', fun=lambda h: np.sum(h) - n_points / 2)],\n options=dict(ftol=1e-8, maxiter=500, disp=False)\n )\n\n best_h = result_slsqp.x\n best_c5 = result_slsqp.fun\n\n return (best_h, best_c5, n_points)\n```",
64 "env/all/time/policy": 527.3870373102836,
65 "env/all/time/policy/min": 231.2769570350647,
66 "env/all/time/policy/max": 755.4097290039062,
67 "env/all/time/env_step": 65.29666293505579,
68 "env/all/time/env_step/min": 0.006223440170288086,
69 "env/all/time/env_step/max": 1100.149849653244,
70 "env/all/time/reward_compute": 3.5064294934272766e-07,
71 "env/all/time/reward_compute/min": 2.7194619178771973e-07,
72 "env/all/time/reward_compute/max": 5.401670932769775e-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": 0.02280581183731556,
77 "advantage/min": -1.0,
78 "advantage/max": 6.029675006866455,
79 "time/assemble_training_data": 8.066887378692627,
80 "time/kl_vs_base": 129.8289394378662,
81 "kl_policy_base": 0.0009261100785806775,
82 "time/train": 1016.4251470565796,
83 "time/save_checkpoint": 18.078874111175537,
84 "time/total": 3039.1853635311127
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