Views
No views yet
erdos. Checkpoint saved
after training step 48 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 48,
3 "progress/batch": 48,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.98,
6 "puct/buffer_size": 770,
7 "puct/sampled_size": 8,
8 "puct/T": 24576,
9 "puct/scale_last": 0.11905941041556217,
10 "puct/buffer_value/mean": -0.3830929569698677,
11 "puct/buffer_value/std": 0.014622503391019736,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809405895844437,
14 "puct/buffer_timestep/mean": 23.11168831168831,
15 "puct/buffer_timestep/std": 13.962305227987024,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 47.0,
18 "puct/buffer_construction_len/mean": 79.54155844155844,
19 "puct/buffer_construction_len/std": 8.546372370084457,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 160.0,
22 "puct/sampled_value/mean": -0.3809407433200901,
23 "puct/sampled_value/std": 9.201706469701165e-08,
24 "puct/sampled_value/min": -0.3809408275716116,
25 "puct/sampled_value/max": -0.3809405895844437,
26 "puct/sampled_timestep/mean": 47.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 47.0,
29 "puct/sampled_timestep/max": 47.0,
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": 3024.6960487365723,
35 "env/all/ac_tokens_per_turn": 9624.966796875,
36 "env/all/ob_tokens_per_turn": 1315.75,
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": 4927983,
41 "env/all/total_ob_tokens": 673664,
42 "env/all/time/sampling_mean": 302.04925864143297,
43 "env/all/time/sampling_max": 428.5966441631317,
44 "env/all/time/env_step_mean": 956.6222157301381,
45 "env/all/time/env_step_max": 2602.456234693527,
46 "env/all/reward/mean": 0.40208826490713623,
47 "env/all/reward/max": 2.6250816718973007,
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": 0.40208826490713623,
53 "env/all/correctness": 0.154296875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.38442190596342685,
57 "env/all/raw_score/min": 0.38094051871019485,
58 "env/all/raw_score/max": 0.5049912602123909,
59 "env/all/initial_raw_score": -0.3809407433200901,
60 "env/all/initial_raw_score/min": -0.3809408275716116,
61 "env/all/initial_raw_score/max": -0.3809405895844437,
62 "env/all/msg": "Success; raw_score=0.3809417857883097",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\n\ndef evaluate_c5(h, dx):\n \"\"\"\n Evaluates the maximum overlap integral C\u2085 for a given step function.\n Now computes the maximum over all valid shifts.\n \"\"\"\n n_points = len(h)\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n max_corr = np.max(corr)\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n\n # Use provided initial guess if available\n if 'initial_h_values' in globals() and len(initial_h_values) > 0:\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n initial_guess = initial_h_values.copy()\n else:\n # Fallback to random binary initialization with exactly half 1s\n n_points = 80\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n # Generate initial guess with random binary vector\n indices = np.random.choice(n_points, int(required_sum), replace=False)\n initial_guess = np.zeros(n_points)\n initial_guess[indices] = 1.0\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 deviation from required sum (moderate penalty)\n penalty = 1e4 * (sum_h - required_sum) ** 2\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Run differential evolution for global exploration\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=50,\n maxiter=600,\n tol=1e-6,\n mutation=(0.4, 0.9),\n recombination=0.9,\n seed=seed,\n x0=initial_guess\n )\n\n best_h = result_de.x\n best_c5 = evaluate_c5(best_h, dx)\n\n # Local refinement with SLSQP and equality constraint\n constraints = [\n {'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum}\n ]\n result_slsqp = minimize(\n objective,\n best_h,\n bounds=bounds,\n method='SLSQP',\n constraints=constraints,\n tol=1e-8,\n options={'maxiter': 500, 'ftol': 1e-5, 'eps': 1e-5}\n )\n\n final_h = result_slsqp.x\n final_c5 = evaluate_c5(final_h, dx)\n\n return final_h, final_c5, n_points\n```",
64 "env/all/time/policy": 302.04925864143297,
65 "env/all/time/policy/min": 120.51657319068909,
66 "env/all/time/policy/max": 428.5966441631317,
67 "env/all/time/env_step": 956.6222157301381,
68 "env/all/time/env_step/min": 0.005092144012451172,
69 "env/all/time/env_step/max": 2602.456234693527,
70 "env/all/time/reward_compute": 2.5797635316848755e-07,
71 "env/all/time/reward_compute/min": 1.341104507446289e-07,
72 "env/all/time/reward_compute/max": 3.948807716369629e-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": 0.055696893483400345,
77 "advantage/min": -1.0,
78 "advantage/max": 27.78572654724121,
79 "time/assemble_training_data": 5.862992763519287,
80 "time/kl_vs_base": 80.8469250202179,
81 "kl_policy_base": 0.0009117589797824621,
82 "time/train": 555.465767621994,
83 "time/save_checkpoint": 10.713650465011597,
84 "time/total": 3679.279440164566
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