Views
No views yet
erdos. Checkpoint saved
after training step 25 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 25,
3 "progress/batch": 25,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.52,
6 "puct/buffer_size": 408,
7 "puct/sampled_size": 8,
8 "puct/T": 12800,
9 "puct/scale_last": 0.1190747377095297,
10 "puct/buffer_value/mean": -0.3845000436856819,
11 "puct/buffer_value/std": 0.018542413682788837,
12 "puct/buffer_value/min": -0.5130522804051018,
13 "puct/buffer_value/max": -0.3809252622904703,
14 "puct/buffer_timestep/mean": 11.745098039215685,
15 "puct/buffer_timestep/std": 7.36404390518193,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 24.0,
18 "puct/buffer_construction_len/mean": 85.67647058823529,
19 "puct/buffer_construction_len/std": 26.241585678380826,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 200.0,
22 "puct/sampled_value/mean": -0.3809252952211442,
23 "puct/sampled_value/std": 1.9080557195415165e-08,
24 "puct/sampled_value/min": -0.38092531369709437,
25 "puct/sampled_value/max": -0.38092526541595606,
26 "puct/sampled_timestep/mean": 24.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 24.0,
29 "puct/sampled_timestep/max": 24.0,
30 "puct/sampled_construction_len/mean": 125.0,
31 "puct/sampled_construction_len/std": 43.30127018922193,
32 "puct/sampled_construction_len/min": 100.0,
33 "puct/sampled_construction_len/max": 200.0,
34 "time/sampling": 3780.363117456436,
35 "env/all/ac_tokens_per_turn": 9165.755859375,
36 "env/all/ob_tokens_per_turn": 1421.5,
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": 4692867,
41 "env/all/total_ob_tokens": 727808,
42 "env/all/time/sampling_mean": 502.95190432993695,
43 "env/all/time/sampling_max": 733.8908505439758,
44 "env/all/time/env_step_mean": 1270.6955336844549,
45 "env/all/time/env_step_max": 3051.6029658317566,
46 "env/all/reward/mean": 0.5691579330294767,
47 "env/all/reward/max": 2.625186823874121,
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": 0.5691579330294767,
53 "env/all/correctness": 0.228515625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.4086719703563173,
57 "env/all/raw_score/min": 0.38092526012010874,
58 "env/all/raw_score/max": 0.9999999999999998,
59 "env/all/initial_raw_score": -0.38092529522114427,
60 "env/all/initial_raw_score/min": -0.38092531369709437,
61 "env/all/initial_raw_score/max": -0.38092526541595606,
62 "env/all/msg": "RuntimeError: Program execution failed: ValueError: operands could not be broadcast together with shapes (100,) (150,) ",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 150 # Larger number for better precision, but within 1k limit\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Initialize with known good starting point\n initial_h = np.array(initial_h_values, dtype=np.float64)\n # Ensure integral constraint\n if not np.allclose(np.sum(initial_h), n_points / 2):\n delta = (n_points / 2) - np.sum(initial_h)\n initial_h[-1] += delta\n initial_h = np.clip(initial_h, 0.0, 1.0)\n\n # Objective function: max overlap, ensuring constraint during optimization\n def objective(h):\n # Enforce integral constraint (adjust last element to meet sum)\n sum_adj = np.sum(h)\n delta = (n_points / 2) - sum_adj\n h_adj = h.copy()\n if abs(delta) > 1e-8:\n h_adj[-1] += delta\n h_adj[-1] = np.clip(h_adj[-1], 0.0, 1.0)\n\n # Compute cross-correlation\n corr = np.correlate(h_adj, 1 - h_adj, mode=\"full\")\n valid_corr = corr[-n_points:] # Only consider k >= 0\n max_overlap = np.max(valid_corr) * dx\n return max_overlap\n\n # Define bounds and constraints\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Run Differential Evolution with constraint enforcement\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n maxiter=200,\n popsize=100,\n mutation=(0.7, 1.0),\n recombination=0.8,\n tol=1e-5,\n seed=seed,\n x0=initial_h\n )\n\n # Refine with SLSQP using equality constraint\n def constraint_func(h):\n return np.sum(h) - n_points / 2\n\n res = minimize(\n fun=objective,\n x0=result_de.x,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={\n 'ftol': 1e-8,\n 'maxiter': 500,\n 'disp': False\n }\n )\n\n return (res.x, res.fun, n_points)\n```",
64 "env/all/time/policy": 502.95190432993695,
65 "env/all/time/policy/min": 247.31238460540771,
66 "env/all/time/policy/max": 733.8908505439758,
67 "env/all/time/env_step": 1270.6955336844549,
68 "env/all/time/env_step/min": 0.0073032379150390625,
69 "env/all/time/env_step/max": 3051.6029658317566,
70 "env/all/time/reward_compute": 4.1816383600234985e-07,
71 "env/all/time/reward_compute/min": 2.682209014892578e-07,
72 "env/all/time/reward_compute/max": 9.648501873016357e-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": 0.03177059069275856,
77 "advantage/min": -0.8564797639846802,
78 "advantage/max": 4.888974666595459,
79 "time/assemble_training_data": 5.958880424499512,
80 "time/kl_vs_base": 127.01605892181396,
81 "kl_policy_base": 0.0008822573581710458,
82 "time/train": 983.0328302383423,
83 "time/save_checkpoint": 18.817044973373413,
84 "time/total": 4916.387994289398
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