Views
No views yet
erdos. Checkpoint saved
after training step 8 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 8,
3 "progress/batch": 8,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.18,
6 "puct/buffer_size": 136,
7 "puct/sampled_size": 8,
8 "puct/T": 4096,
9 "puct/scale_last": 0.0676615381901251,
10 "puct/buffer_value/mean": -0.39026032383078285,
11 "puct/buffer_value/std": 0.029981822248180214,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809428882921597,
14 "puct/buffer_timestep/mean": 3.235294117647059,
15 "puct/buffer_timestep/std": 2.462170533700747,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 7.0,
18 "puct/buffer_construction_len/mean": 76.5220588235294,
19 "puct/buffer_construction_len/std": 18.7000184459506,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.3809563414275423,
23 "puct/sampled_value/std": 8.976787940398085e-06,
24 "puct/sampled_value/min": -0.3809674698712975,
25 "puct/sampled_value/max": -0.3809428882921597,
26 "puct/sampled_timestep/mean": 7.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 7.0,
29 "puct/sampled_timestep/max": 7.0,
30 "puct/sampled_construction_len/mean": 80.0,
31 "puct/sampled_construction_len/std": 0.0,
32 "puct/sampled_construction_len/min": 80.0,
33 "puct/sampled_construction_len/max": 80.0,
34 "time/sampling": 1981.835171699524,
35 "env/all/ac_tokens_per_turn": 9338.53125,
36 "env/all/ob_tokens_per_turn": 1343.5,
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": 4781328,
41 "env/all/total_ob_tokens": 687872,
42 "env/all/time/sampling_mean": 500.7550093368627,
43 "env/all/time/sampling_max": 726.5329070091248,
44 "env/all/time/env_step_mean": 151.6333510922268,
45 "env/all/time/env_step_max": 1254.8406417369843,
46 "env/all/reward/mean": 0.8763558373737096,
47 "env/all/reward/max": 2.625072731298896,
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": 0.8763558373737096,
53 "env/all/correctness": 0.353515625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.4092054744429947,
57 "env/all/raw_score/min": 0.38094181613568817,
58 "env/all/raw_score/max": 0.8245412421458769,
59 "env/all/initial_raw_score": -0.3809563414275424,
60 "env/all/initial_raw_score/min": -0.3809674698712975,
61 "env/all/initial_raw_score/max": -0.3809428882921597,
62 "env/all/msg": "Success; raw_score=0.3826871401322771",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, Bounds, minimize\nfrom scipy.signal import correlate\n\ndef evaluate_C5(h, dx):\n \"\"\"Compute the maximum overlap integral with 1 - h over all shifts.\"\"\"\n h1 = 1.0 - h\n corr = correlate(h, h1, mode='full')\n max_corr = np.max(corr)\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n import random\n random.seed(seed)\n np.random.seed(seed)\n\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = int(n_points / 2)\n\n # Define the objective function with normalization\n def objective(h_vec):\n h = np.array(h_vec)\n if np.sum(h) == 0:\n return float('inf')\n scale = required_sum / np.sum(h)\n h_scaled = h * scale\n return evaluate_C5(h_scaled, dx)\n\n # Define bounds (each h_i \u2208 [0, 1])\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Generate diverse initial guesses\n def generate_initial_guesses(n_points, required_sum):\n guesses = []\n # 3 structured guesses\n for shift in range(3):\n h = np.zeros(n_points)\n for i in range(n_points):\n if i % (n_points // 2) == shift:\n h[i] = 1\n guesses.append(h)\n # 2 random guesses\n for _ in range(2):\n h = np.random.rand(n_points)\n h = h / h.sum() * required_sum\n guesses.append(h)\n return guesses\n\n initial_guesses = generate_initial_guesses(n_points, required_sum)\n\n best_h = initial_h_values.copy()\n best_c5 = float('inf')\n\n # Run Differential Evolution\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=15,\n maxiter=100,\n tol=1e-8,\n mutation=(0.7, 1.0),\n recombination=0.8\n )\n\n # Refine with SLSQP\n h_de = result_de.x\n cons = {'type': 'eq', 'fun': lambda h: np.sum(h) - required_sum}\n local_bounds = Bounds(0.0, 1.0)\n\n result_slsqp = minimize(\n fun=objective,\n x0=h_de,\n method='SLSQP',\n bounds=local_bounds,\n constraints=cons,\n options={'ftol': 1e-9, 'maxiter': 300}\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": 500.7550093368627,
65 "env/all/time/policy/min": 208.38646173477173,
66 "env/all/time/policy/max": 726.5329070091248,
67 "env/all/time/env_step": 151.6333510922268,
68 "env/all/time/env_step/min": 0.0060465335845947266,
69 "env/all/time/env_step/max": 1254.8406417369843,
70 "env/all/time/reward_compute": 4.507601261138916e-07,
71 "env/all/time/reward_compute/min": 3.427267074584961e-07,
72 "env/all/time/reward_compute/max": 8.083879947662354e-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": 0.02529784105718136,
77 "advantage/min": -0.9723279476165771,
78 "advantage/max": 4.754220962524414,
79 "time/assemble_training_data": 9.860372543334961,
80 "time/kl_vs_base": 125.49159073829651,
81 "kl_policy_base": 0.0007020130869932473,
82 "time/train": 997.8491721153259,
83 "time/save_checkpoint": 12.710592985153198,
84 "time/total": 3129.018468141556
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