Views
No views yet
erdos. Checkpoint saved
after training step 15 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 15,
3 "progress/batch": 15,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.32,
6 "puct/buffer_size": 248,
7 "puct/sampled_size": 8,
8 "puct/T": 7680,
9 "puct/scale_last": 0.11905839782286143,
10 "puct/buffer_value/mean": -0.3865490919261261,
11 "puct/buffer_value/std": 0.02379734936195881,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38094160217714446,
14 "puct/buffer_timestep/mean": 6.741935483870968,
15 "puct/buffer_timestep/std": 4.479110956828331,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 14.0,
18 "puct/buffer_construction_len/mean": 78.17338709677419,
19 "puct/buffer_construction_len/std": 14.024107175715773,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38094162184773095,
23 "puct/sampled_value/std": 1.476946760024124e-08,
24 "puct/sampled_value/min": -0.38094164758418836,
25 "puct/sampled_value/max": -0.38094160217714446,
26 "puct/sampled_timestep/mean": 14.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 14.0,
29 "puct/sampled_timestep/max": 14.0,
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": 1758.1136813163757,
35 "env/all/ac_tokens_per_turn": 9349.978515625,
36 "env/all/ob_tokens_per_turn": 1633.25,
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": 4787189,
41 "env/all/total_ob_tokens": 836224,
42 "env/all/time/sampling_mean": 511.4282429860905,
43 "env/all/time/sampling_max": 767.5150606632233,
44 "env/all/time/env_step_mean": 30.659455427899957,
45 "env/all/time/env_step_max": 1100.1287834644318,
46 "env/all/reward/mean": 0.9675884296834504,
47 "env/all/reward/max": 2.6250745186390128,
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": 0.9675884296834504,
53 "env/all/correctness": 0.384765625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.40175073618628804,
57 "env/all/raw_score/min": 0.38094155676301006,
58 "env/all/raw_score/max": 0.5183051091884735,
59 "env/all/initial_raw_score": -0.38094162184773084,
60 "env/all/initial_raw_score/min": -0.38094164758418836,
61 "env/all/initial_raw_score/max": -0.38094160217714446,
62 "env/all/msg": "RuntimeError: Program execution failed: ValueError: The population supplied needs to have shape (S, len(x)), where S > 4.",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\nfrom scipy.signal import correlate\nimport time\n\ndef evaluate_C5(h, dx):\n h1 = 1.0 - h\n corr = 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_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = int(n_points / 2)\n\n start = time.time()\n\n # Objective function with constraint enforcement\n def objective(x):\n current_sum = np.sum(x)\n scaling_factor = required_sum / current_sum\n if current_sum > required_sum:\n adjusted_x = x * scaling_factor\n return evaluate_C5(adjusted_x, dx)\n elif current_sum < required_sum:\n # Penalize for sum below required\n return evaluate_C5(x, dx) + 1000 * (required_sum - current_sum)\n else:\n return evaluate_C5(x, dx)\n\n # Generate initial guess based on the provided seed\n h_initial = initial_h_values.copy()\n h_initial = h_initial * (required_sum / h_initial.sum())\n\n # Bounds for each parameter\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Run Differential Evolution to explore the search space\n de_result = differential_evolution(\n objective,\n bounds,\n popsize=100, # Higher population size for better exploration\n mutation=(0.8, 1.2),\n maxiter=500, # Reduced to save time\n tol=1e-6,\n polish=True,\n init=[h_initial],\n strategy='best1bin'\n )\n\n # Local optimization with SLSQP, ensuring sum constraint is satisfied\n def objective_local(x):\n return evaluate_C5(x, dx)\n\n constraint = {'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum}\n\n # Run multiple local optimizations to avoid getting stuck in poor local optima\n num_local_optimizations = 5\n local_optimizers = []\n\n for _ in range(num_local_optimizations):\n result_local = minimize(\n objective_local,\n de_result.x,\n method='SLSQP',\n bounds=bounds,\n constraints=constraint,\n options={'maxiter': 300, 'ftol': 1e-9, 'disp': False}\n )\n local_optimizers.append(result_local)\n\n # Select the best among the local optimizations\n best_result = min(local_optimizers, key=lambda r: r.fun)\n\n # Ensure the final solution meets the constraint\n best_h = best_result.x\n best_c5 = best_result.fun\n\n elapsed = time.time() - start\n if elapsed > budget_s:\n print(f\"Budget exceeded: {elapsed} seconds used.\")\n return de_result.x, de_result.fun, n_points # Fallback\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 511.4282429860905,
65 "env/all/time/policy/min": 192.22372603416443,
66 "env/all/time/policy/max": 767.5150606632233,
67 "env/all/time/env_step": 30.659455427899957,
68 "env/all/time/env_step/min": 0.005598306655883789,
69 "env/all/time/env_step/max": 1100.1287834644318,
70 "env/all/time/reward_compute": 2.9243528842926025e-07,
71 "env/all/time/reward_compute/min": 2.123415470123291e-07,
72 "env/all/time/reward_compute/max": 4.0605664253234863e-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": 0.0224761925637722,
77 "advantage/min": -0.9922894835472107,
78 "advantage/max": 2.3455491065979004,
79 "time/assemble_training_data": 9.034083604812622,
80 "time/kl_vs_base": 129.9757800102234,
81 "kl_policy_base": 0.0008436993230134249,
82 "time/train": 1026.899269104004,
83 "time/save_checkpoint": 9.124225378036499,
84 "time/total": 2934.2465369701385
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