Views
No views yet
erdos. Checkpoint saved
after training step 22 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 22,
3 "progress/batch": 22,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.46,
6 "puct/buffer_size": 360,
7 "puct/sampled_size": 8,
8 "puct/T": 11264,
9 "puct/scale_last": 0.11905844757870132,
10 "puct/buffer_value/mean": -0.38487862503745923,
11 "puct/buffer_value/std": 0.019944816391693834,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38094155242130456,
14 "puct/buffer_timestep/mean": 10.244444444444444,
15 "puct/buffer_timestep/std": 6.4983948635389295,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 21.0,
18 "puct/buffer_construction_len/mean": 78.74166666666666,
19 "puct/buffer_construction_len/std": 11.670591601685553,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38094155441453853,
23 "puct/sampled_value/std": 1.2774584847050196e-09,
24 "puct/sampled_value/min": -0.3809415560130435,
25 "puct/sampled_value/max": -0.38094155242130456,
26 "puct/sampled_timestep/mean": 21.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 21.0,
29 "puct/sampled_timestep/max": 21.0,
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": 3118.260138988495,
35 "env/all/ac_tokens_per_turn": 9316.9296875,
36 "env/all/ob_tokens_per_turn": 1419.375,
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": 4770268,
41 "env/all/total_ob_tokens": 726720,
42 "env/all/time/sampling_mean": 504.79024944407865,
43 "env/all/time/sampling_max": 731.4294803142548,
44 "env/all/time/env_step_mean": 999.9643861451186,
45 "env/all/time/env_step_max": 2415.6749238967896,
46 "env/all/reward/mean": 0.6440495639281556,
47 "env/all/reward/max": 2.6250745852928548,
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": 0.6440495639281556,
53 "env/all/correctness": 0.251953125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.39296026915475923,
57 "env/all/raw_score/min": 0.3809415470904396,
58 "env/all/raw_score/max": 0.5162936382066062,
59 "env/all/initial_raw_score": -0.38094155441453853,
60 "env/all/initial_raw_score/min": -0.3809415560130435,
61 "env/all/initial_raw_score/max": -0.38094155242130456,
62 "env/all/msg": "After normalization, h(x) is not in [0, 1]. Range: [0.016254745639598025, 1.0000000386686454]",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\nfrom scipy.signal import correlate\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n \n # Assume global variable initial_h_values is defined\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n\n # Generate a structured initial guess: alternating 1s and 0s\n block_size = n_points // 4\n structured_h = np.zeros(n_points)\n for i in range(n_points):\n block = i // block_size\n if block % 2 == 0: # even blocks\n structured_h[i] = 1.0\n else:\n structured_h[i] = 0.0\n\n # Use the structured guess as initial guess\n initial_guess = structured_h.copy()\n\n def objective(x):\n h = x\n h1 = 1.0 - h\n corr = correlate(h, h1, mode='full')\n max_corr = np.max(corr)\n # Enforce constraint with a strong penalty\n penalty = 1e5 * (np.sum(h) - required_sum) ** 2\n return max_corr * dx + penalty\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Run differential evolution with optimized parameters\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=60,\n maxiter=1000, # Reduced from 1500 to stay within budget\n tol=1e-5,\n mutation=(0.8, 1.0),\n recombination=0.9,\n disp=False,\n polish=True,\n x0=initial_guess\n )\n\n best_h = result.x\n h = best_h\n h1 = 1.0 - h\n corr = correlate(h, h1, mode='full')\n max_corr = np.max(corr)\n true_c5 = max_corr * dx\n\n return best_h, true_c5, n_points\n```",
64 "env/all/time/policy": 504.79024944407865,
65 "env/all/time/policy/min": 227.19003582000732,
66 "env/all/time/policy/max": 731.4294803142548,
67 "env/all/time/env_step": 999.9643861451186,
68 "env/all/time/env_step/min": 0.005814075469970703,
69 "env/all/time/env_step/max": 2415.6749238967896,
70 "env/all/time/reward_compute": 3.5855919122695923e-07,
71 "env/all/time/reward_compute/min": 2.3096799850463867e-07,
72 "env/all/time/reward_compute/max": 5.736947059631348e-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": 0.03102775663137436,
77 "advantage/min": -0.9537604451179504,
78 "advantage/max": 6.081274509429932,
79 "time/assemble_training_data": 7.971552133560181,
80 "time/kl_vs_base": 134.35003352165222,
81 "kl_policy_base": 0.0008775901515036821,
82 "time/train": 1000.8352000713348,
83 "time/save_checkpoint": 11.143954992294312,
84 "time/total": 4273.980872631073
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