Views
No views yet
erdos. Checkpoint saved
after training step 38 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 38,
3 "progress/batch": 38,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.78,
6 "puct/buffer_size": 615,
7 "puct/sampled_size": 8,
8 "puct/T": 19456,
9 "puct/scale_last": 0.1190591760561926,
10 "puct/buffer_value/mean": -0.3832540414224449,
11 "puct/buffer_value/std": 0.015381846928008442,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809408239438133,
14 "puct/buffer_timestep/mean": 18.227642276422763,
15 "puct/buffer_timestep/std": 11.114783978846557,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 37.0,
18 "puct/buffer_construction_len/mean": 79.26341463414634,
19 "puct/buffer_construction_len/std": 8.950571297501392,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.3809412553258566,
23 "puct/sampled_value/std": 2.405867176982154e-07,
24 "puct/sampled_value/min": -0.3809414695474924,
25 "puct/sampled_value/max": -0.38094085726727944,
26 "puct/sampled_timestep/mean": 37.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 37.0,
29 "puct/sampled_timestep/max": 37.0,
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": 4418.391325950623,
35 "env/all/ac_tokens_per_turn": 9223.052734375,
36 "env/all/ob_tokens_per_turn": 1210.5,
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": 4722203,
41 "env/all/total_ob_tokens": 619776,
42 "env/all/time/sampling_mean": 275.9215144198388,
43 "env/all/time/sampling_max": 391.3972408771515,
44 "env/all/time/env_step_mean": 1937.9432481480762,
45 "env/all/time/env_step_max": 4033.250534772873,
46 "env/all/reward/mean": 0.8198317265742484,
47 "env/all/reward/max": 2.625079420678465,
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": 0.8198317265742484,
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": 0.3840582880372554,
57 "env/all/raw_score/min": 0.38094084539764167,
58 "env/all/raw_score/max": 0.4901738130349654,
59 "env/all/initial_raw_score": -0.38094125532585665,
60 "env/all/initial_raw_score/min": -0.3809414695474924,
61 "env/all/initial_raw_score/max": -0.38094085726727944,
62 "env/all/msg": "After normalization, h(x) is not in [0, 1]. Range: [0.0, 1.3240218803332264]",
63 "env/all/parsed_code": "```python\nimport numpy as np\nimport time\nfrom scipy.optimize import differential_evolution, minimize\n\ndef evaluate_c5(h_values, dx):\n \"\"\"Evaluate the maximum overlap integral for a step function h_values.\"\"\"\n one_minus_h = 1.0 - h_values\n corr = np.correlate(h_values, one_minus_h, mode='full')\n # Focus on the central portion to capture relevant shifts\n max_corr = np.max(corr[len(h_values) - 1 : 2 * len(h_values) - 1])\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n import random\n random.seed(seed)\n\n # Use the initial_h_values if provided\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2\n\n # Use an initial guess with randomized values if no initial_h_values provided\n if initial_h_values is None:\n initial_guess = np.random.rand(n_points)\n initial_guess = initial_guess / np.mean(initial_guess) * required_sum\n else:\n initial_guess = initial_h_values.copy()\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(x):\n h = np.array(x)\n sum_h = np.sum(h)\n # Penalize constraint violation with a lower coefficient\n penalty = 1e3 * (sum_h - required_sum) ** 2\n c5 = evaluate_c5(h, dx)\n return c5 + penalty\n\n # Run Differential Evolution\n start_time = time.time()\n result = differential_evolution(\n objective,\n bounds,\n strategy='rand1bin', # Better balance between exploration and improvement\n popsize=60, # Larger population for diversity\n maxiter=1500, # More iterations for refinement\n tol=1e-6,\n mutation=(0.5, 0.8), # Adjusted mutation range\n recombination=0.8,\n seed=seed,\n x0=initial_guess\n )\n\n best_h = result.x\n best_c5 = evaluate_c5(best_h, dx)\n\n # Local refinement using Nelder-Mead\n # We refine the best found solution\n res = minimize(\n lambda x: evaluate_c5(x, dx),\n best_h,\n method='Nelder-Mead',\n bounds=bounds\n )\n\n best_h = res.x\n best_c5 = evaluate_c5(best_h, dx)\n\n elapsed_time = time.time() - start_time\n if elapsed_time > budget_s:\n print(\"Warning: Optimization took more than the budget time.\")\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 275.9215144198388,
65 "env/all/time/policy/min": 128.58829641342163,
66 "env/all/time/policy/max": 391.3972408771515,
67 "env/all/time/env_step": 1937.9432481480762,
68 "env/all/time/env_step/min": 0.005029439926147461,
69 "env/all/time/env_step/max": 4033.250534772873,
70 "env/all/time/reward_compute": 3.394670784473419e-07,
71 "env/all/time/reward_compute/min": 2.3469328880310059e-07,
72 "env/all/time/reward_compute/max": 5.550682544708252e-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": 0.037511035799980164,
77 "advantage/min": -1.0,
78 "advantage/max": 27.785701751708984,
79 "time/assemble_training_data": 9.13228440284729,
80 "time/kl_vs_base": 78.78413820266724,
81 "kl_policy_base": 0.0009246038389392197,
82 "time/train": 523.4017133712769,
83 "time/save_checkpoint": 14.2970712184906,
84 "time/total": 5045.4745881557465
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
[2026-07-10T08:20:19+00:00] job=1813610 node=node-29 ngpu=6 ntrain=2 replicas=4 flash_attn=yes