Views
No views yet
erdos. Checkpoint saved
after training step 24 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 24,
3 "progress/batch": 24,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.5,
6 "puct/buffer_size": 392,
7 "puct/sampled_size": 8,
8 "puct/T": 12288,
9 "puct/scale_last": 0.11905846516654633,
10 "puct/buffer_value/mean": -0.3845572313150543,
11 "puct/buffer_value/std": 0.01914378852556912,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38094153483345955,
14 "puct/buffer_timestep/mean": 11.244897959183673,
15 "puct/buffer_timestep/std": 7.075484009534579,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 23.0,
18 "puct/buffer_construction_len/mean": 78.84438775510205,
19 "puct/buffer_construction_len/std": 11.189406849586103,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38094154583372997,
23 "puct/sampled_value/std": 4.725430310130379e-09,
24 "puct/sampled_value/min": -0.38094155157723597,
25 "puct/sampled_value/max": -0.38094153483345955,
26 "puct/sampled_timestep/mean": 23.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 23.0,
29 "puct/sampled_timestep/max": 23.0,
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": 4514.420679330826,
35 "env/all/ac_tokens_per_turn": 9385.19921875,
36 "env/all/ob_tokens_per_turn": 1365.625,
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": 4805222,
41 "env/all/total_ob_tokens": 699200,
42 "env/all/time/sampling_mean": 513.68874246208,
43 "env/all/time/sampling_max": 747.7559804916382,
44 "env/all/time/env_step_mean": 1768.4973045126535,
45 "env/all/time/env_step_max": 3765.464462995529,
46 "env/all/reward/mean": 0.33370990319894817,
47 "env/all/reward/max": 2.6250747054422106,
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": 0.33370990319894817,
53 "env/all/correctness": 0.12890625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.3874785092997671,
57 "env/all/raw_score/min": 0.38094152965479,
58 "env/all/raw_score/max": 0.5,
59 "env/all/initial_raw_score": -0.38094154583373013,
60 "env/all/initial_raw_score/min": -0.38094155157723597,
61 "env/all/initial_raw_score/max": -0.38094153483345955,
62 "env/all/msg": "RuntimeError: Program execution failed: ValueError: operands could not be broadcast together with shapes (80,) (60,) ",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\n\ndef evaluate_C5(h, dx):\n h1 = 1.0 - h\n corr = np.correlate(h, h1, 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 n_points = 60\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n\n def objective_for_de(h):\n c5 = evaluate_C5(h, dx)\n sum_penalty = 100.0 * (np.sum(h) - required_sum) ** 2 # Increased penalty\n return c5 + sum_penalty\n\n def objective_for_local(h):\n return evaluate_C5(h, dx)\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n initial_guess = initial_h_values.copy()\n\n result_de = differential_evolution(\n objective_for_de,\n bounds,\n strategy='rand1bin', # Changed from 'best1bin'\n popsize=100,\n maxiter=500, # Reduced iterations for efficiency\n mutation=(0.5, 1.5), # Tuned mutation range\n recombination=0.9, # Increased recombination\n tol=1e-6,\n polish=True,\n x0=initial_guess,\n disp=False\n )\n\n best_h_de = result_de.x\n\n # Local optimization with SLSQP and equality constraint\n result_local = minimize(\n objective_for_local,\n best_h_de,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum}],\n tol=1e-6,\n options={'maxiter': 200, 'disp': False}\n )\n\n best_h = result_local.x\n best_c5 = evaluate_C5(best_h, dx)\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 513.68874246208,
65 "env/all/time/policy/min": 211.47952318191528,
66 "env/all/time/policy/max": 747.7559804916382,
67 "env/all/time/env_step": 1768.4973045126535,
68 "env/all/time/env_step/min": 0.0062580108642578125,
69 "env/all/time/env_step/max": 3765.464462995529,
70 "env/all/time/reward_compute": 7.557682693004608e-07,
71 "env/all/time/reward_compute/min": 2.905726432800293e-07,
72 "env/all/time/reward_compute/max": 3.5315752029418945e-06,
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": 0.043530020862817764,
77 "advantage/min": -0.6881158351898193,
78 "advantage/max": 14.112600326538086,
79 "time/assemble_training_data": 7.767157316207886,
80 "time/kl_vs_base": 129.4795322418213,
81 "kl_policy_base": 0.0009028607164509594,
82 "time/train": 1004.811418056488,
83 "time/save_checkpoint": 3.2406625747680664,
84 "time/total": 5661.513519287109
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