Views
No views yet
erdos. Checkpoint saved
after training step 26 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 26,
3 "progress/batch": 26,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.54,
6 "puct/buffer_size": 424,
7 "puct/sampled_size": 8,
8 "puct/T": 13312,
9 "puct/scale_last": 0.119058473847169,
10 "puct/buffer_value/mean": -0.3842894072646515,
11 "puct/buffer_value/std": 0.01843134863978844,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809415261528369,
14 "puct/buffer_timestep/mean": 12.245283018867925,
15 "puct/buffer_timestep/std": 7.652612478761028,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 25.0,
18 "puct/buffer_construction_len/mean": 78.93160377358491,
19 "puct/buffer_construction_len/std": 10.763212208748456,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.3809415394232218,
23 "puct/sampled_value/std": 5.242277041514456e-09,
24 "puct/sampled_value/min": -0.38094154299314054,
25 "puct/sampled_value/max": -0.3809415261528369,
26 "puct/sampled_timestep/mean": 25.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 25.0,
29 "puct/sampled_timestep/max": 25.0,
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": 4024.556714296341,
35 "env/all/ac_tokens_per_turn": 9789.8203125,
36 "env/all/ob_tokens_per_turn": 1508.5,
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": 5012388,
41 "env/all/total_ob_tokens": 772352,
42 "env/all/time/sampling_mean": 317.180385726504,
43 "env/all/time/sampling_max": 467.47024631500244,
44 "env/all/time/env_step_mean": 1565.526480378583,
45 "env/all/time/env_step_max": 3588.048487186432,
46 "env/all/reward/mean": 0.5769741767609853,
47 "env/all/reward/max": 2.6250755793090317,
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": 0.5769741767609853,
53 "env/all/correctness": 0.220703125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.38283680648413626,
57 "env/all/raw_score/min": 0.38094140284237554,
58 "env/all/raw_score/max": 0.5,
59 "env/all/initial_raw_score": -0.3809415394232218,
60 "env/all/initial_raw_score/min": -0.38094154299314054,
61 "env/all/initial_raw_score/max": -0.3809415261528369,
62 "env/all/msg": "Success; raw_score=0.3809416639524406",
63 "env/all/parsed_code": "```python\nimport numpy as np\nimport time\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2.0 # sum(h) * dx = 1 => sum(h) = required_sum\n\n # Initial solution\n current_h = initial_h_values.copy()\n best_h = current_h.copy()\n current_c5 = compute_c5(current_h, n_points, dx)\n best_c5 = current_c5\n\n # Annealing parameters\n temperature = 1.0\n cooling_rate = 0.9995\n max_iterations = int(budget_s * 1000) # roughly 1000 seconds with 1000 iterations per second\n\n start_time = time.time()\n\n for iteration in range(max_iterations):\n if time.time() - start_time > budget_s:\n break\n\n # Perturbation step\n i, j = np.random.choice(n_points, size=2, replace=False)\n delta = np.random.uniform(-0.05, 0.05)\n\n old_h_i = current_h[i]\n new_h_i = np.clip(old_h_i + delta, 0.0, 1.0)\n delta_sum = new_h_i - old_h_i\n\n old_h_j = current_h[j]\n new_h_j = np.clip(old_h_j - delta_sum, 0.0, 1.0)\n\n new_h = current_h.copy()\n new_h[i] = new_h_i\n new_h[j] = new_h_j\n\n # Evaluate new solution\n new_c5 = compute_c5(new_h, n_points, dx)\n delta_c5 = new_c5 - current_c5\n\n # Acceptance probability\n if delta_c5 < 0:\n current_h = new_h\n current_c5 = new_c5\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_h = current_h.copy()\n else:\n probability = np.exp(-delta_c5 / temperature)\n if np.random.rand() < probability:\n current_h = new_h\n current_c5 = new_c5\n\n # Cool down\n temperature *= cooling_rate\n\n return best_h, best_c5, n_points\n\ndef compute_c5(h, n_points, dx):\n crosscorr = np.correlate(h, 1 - h, mode='full')\n return np.max(crosscorr) * dx\n```",
64 "env/all/time/policy": 317.180385726504,
65 "env/all/time/policy/min": 137.86819005012512,
66 "env/all/time/policy/max": 467.47024631500244,
67 "env/all/time/env_step": 1565.526480378583,
68 "env/all/time/env_step/min": 0.005923271179199219,
69 "env/all/time/env_step/max": 3588.048487186432,
70 "env/all/time/reward_compute": 3.473833203315735e-07,
71 "env/all/time/reward_compute/min": 2.7939677238464355e-07,
72 "env/all/time/reward_compute/max": 6.444752216339111e-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": 0.035595446825027466,
77 "advantage/min": -0.9990447759628296,
78 "advantage/max": 9.353395462036133,
79 "time/assemble_training_data": 8.278056383132935,
80 "time/kl_vs_base": 84.04786348342896,
81 "kl_policy_base": 0.0008570641512051225,
82 "time/train": 577.8851940631866,
83 "time/save_checkpoint": 17.615535974502563,
84 "time/total": 4713.675470590591
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