Views
No views yet
erdos. Checkpoint saved
after training step 47 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 47,
3 "progress/batch": 47,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.96,
6 "puct/buffer_size": 760,
7 "puct/sampled_size": 8,
8 "puct/T": 24064,
9 "puct/scale_last": 0.2791161009493792,
10 "puct/buffer_value/mean": -0.3841118164196077,
11 "puct/buffer_value/std": 0.019073379068698295,
12 "puct/buffer_value/min": -0.66,
13 "puct/buffer_value/max": -0.3808838990506208,
14 "puct/buffer_timestep/mean": 22.74736842105263,
15 "puct/buffer_timestep/std": 13.7135878994081,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 46.0,
18 "puct/buffer_construction_len/mean": 110.58684210526316,
19 "puct/buffer_construction_len/std": 53.7487119603449,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 400.0,
22 "puct/sampled_value/mean": -0.3808847684977377,
23 "puct/sampled_value/std": 5.368714229430111e-07,
24 "puct/sampled_value/min": -0.3808853514332606,
25 "puct/sampled_value/max": -0.3808838990506208,
26 "puct/sampled_timestep/mean": 46.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 46.0,
29 "puct/sampled_timestep/max": 46.0,
30 "puct/sampled_construction_len/mean": 200.0,
31 "puct/sampled_construction_len/std": 0.0,
32 "puct/sampled_construction_len/min": 200.0,
33 "puct/sampled_construction_len/max": 200.0,
34 "time/sampling": 2441.3962075710297,
35 "env/all/ac_tokens_per_turn": 10100.82421875,
36 "env/all/ob_tokens_per_turn": 1802.375,
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": 5171622,
41 "env/all/total_ob_tokens": 922816,
42 "env/all/time/sampling_mean": 591.0793969719671,
43 "env/all/time/sampling_max": 875.8848466873169,
44 "env/all/time/env_step_mean": 299.2029542317614,
45 "env/all/time/env_step_max": 1569.9851822853088,
46 "env/all/reward/mean": 0.6620737486479191,
47 "env/all/reward/max": 2.625471899017418,
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": 0.6620737486479191,
53 "env/all/correctness": 0.2734375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.4207692209215659,
57 "env/all/raw_score/min": 0.3808838990505024,
58 "env/all/raw_score/max": 0.6983621736693861,
59 "env/all/initial_raw_score": -0.3808847684977377,
60 "env/all/initial_raw_score/min": -0.3808853514332606,
61 "env/all/initial_raw_score/max": -0.3808838990506208,
62 "env/all/msg": "After normalization, h(x) is not in [0, 1]. Range: [0.0, 1.0000000000000302]",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n \n n_points = 200\n dx = 2.0 / n_points\n desired_sum = n_points / 2.0\n\n # Start from the current best initial guess\n initial_h = None\n if 'initial_h_values' in globals():\n initial_h = np.array(initial_h_values)\n # Ensure length matches n_points by padding or downscaling\n if len(initial_h) > n_points:\n initial_h = initial_h[::2]\n elif len(initial_h) < n_points:\n period = len(initial_h)\n num_repeats = n_points // period\n initial_h = np.repeat(initial_h, num_repeats)[:n_points]\n # Adjust to meet desired sum\n sum_h = np.sum(initial_h)\n delta = desired_sum - sum_h\n for i in reversed(range(n_points)):\n if delta <= 0:\n break\n add = min(1.0 - initial_h[i], delta)\n initial_h[i] += add\n delta -= add\n initial_h = np.clip(initial_h, 0, 1)\n else:\n initial_h = np.zeros(n_points)\n for i in range(n_points):\n if i % 2 == 0:\n initial_h[i] = 1.0\n\n # Define the objective function\n def objective(h_vec):\n # Compute overlaps for relevant shifts (e.g., k = 0, 0.5, 1, 1.5)\n cross_corr = np.correlate(h_vec, 1 - h_vec, mode='full')\n overlaps = cross_corr * dx\n # We select the top 5 highest overlap values\n selected_overlaps = overlaps[::5]\n max_overlap = np.max(selected_overlaps)\n\n # Penalize deviation from the desired sum\n sum_h = np.sum(h_vec)\n penalty = 1e6 * (sum_h - desired_sum) ** 2\n\n return max_overlap + penalty\n\n # Optimize using differential evolution\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=15,\n mutation=(0.5, 1.0),\n recombination=0.8,\n tol=1e-8,\n maxiter=400,\n disp=False,\n x0=initial_h # try to use the initial guess as starting point\n )\n\n # Return the best solution\n best_h = result.x\n # Recompute the max overlap for accurate result\n cross_corr = np.correlate(best_h, 1 - best_h, mode='full')\n overlaps = cross_corr * dx\n best_c5 = np.max(overlaps)\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 591.0793969719671,
65 "env/all/time/policy/min": 251.44168734550476,
66 "env/all/time/policy/max": 875.8848466873169,
67 "env/all/time/env_step": 299.2029542317614,
68 "env/all/time/env_step/min": 0.007551908493041992,
69 "env/all/time/env_step/max": 1569.9851822853088,
70 "env/all/time/reward_compute": 3.25031578540802e-07,
71 "env/all/time/reward_compute/min": 2.2351741790771484e-07,
72 "env/all/time/reward_compute/max": 4.6938657760620117e-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": 0.029319100081920624,
77 "advantage/min": -0.9437234997749329,
78 "advantage/max": 4.046933174133301,
79 "time/assemble_training_data": 7.115309000015259,
80 "time/kl_vs_base": 137.5044240951538,
81 "kl_policy_base": 0.0008854613406583667,
82 "time/train": 1134.6580882072449,
83 "time/save_checkpoint": 11.1017165184021,
84 "time/total": 3733.3836040496826
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