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": 0.11905847034521588,
10 "puct/buffer_value/mean": -0.384415440130347,
11 "puct/buffer_value/std": 0.018777786307534163,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38094152965479,
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": 78.88970588235294,
19 "puct/buffer_construction_len/std": 10.97010607808672,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.3809415410357246,
23 "puct/sampled_value/std": 5.526188623675067e-09,
24 "puct/sampled_value/min": -0.38094154567519484,
25 "puct/sampled_value/max": -0.38094152965479,
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": 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": 4873.7072496414185,
35 "env/all/ac_tokens_per_turn": 9432.859375,
36 "env/all/ob_tokens_per_turn": 1452.75,
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": 4829624,
41 "env/all/total_ob_tokens": 743808,
42 "env/all/time/sampling_mean": 306.81829041196033,
43 "env/all/time/sampling_max": 409.4239983558655,
44 "env/all/time/env_step_mean": 2136.9269451126456,
45 "env/all/time/env_step_max": 4452.141984701157,
46 "env/all/reward/mean": 0.43708932444530857,
47 "env/all/reward/max": 2.62507472957423,
48 "env/all/reward/min": 0.0,
49 "env/all/format": 0.998046875,
50 "env/all/format/min": 0.0,
51 "env/all/format/max": 1.0,
52 "env/all/reward": 0.43708932444530857,
53 "env/all/correctness": 0.173828125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.4018410817393752,
57 "env/all/raw_score/min": 0.3809415261528369,
58 "env/all/raw_score/max": 0.5,
59 "env/all/initial_raw_score": -0.3809415410357246,
60 "env/all/initial_raw_score/min": -0.38094154567519484,
61 "env/all/initial_raw_score/max": -0.38094152965479,
62 "env/all/msg": "Success; raw_score=0.4290371383937213",
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 generate_random_initial(n_points):\n while True:\n h = np.random.uniform(0, 1, n_points)\n current_sum = np.sum(h)\n required_sum = n_points / 2.0\n if current_sum == 0:\n continue\n scaling = required_sum / current_sum\n scaled_h = h * scaling\n if np.all(scaled_h >= 0) and np.all(scaled_h <= 1):\n return scaled_h\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n \n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n\n # Step 1: Generate better initial guess via random sampling\n initial_guess = generate_random_initial(n_points)\n\n # Objective function for differential evolution\n def objective_for_de(h):\n c5 = evaluate_C5(h, dx)\n return c5\n\n # DE bounds and parameters\n bounds_de = [(0.0, 1.0) for _ in range(n_points)]\n result_de = differential_evolution(\n objective_for_de,\n bounds_de,\n strategy='rand1bin',\n popsize=50, # Balanced size for diversity\n maxiter=1000, # Extended for thorough search\n mutation=(0.9, 1.0), # Wide mutation range for exploration\n recombination=0.95,\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 # Define objective and constraints for local optimization\n def objective(h):\n return evaluate_C5(h, dx)\n\n def constraint(h):\n return np.sum(h) - required_sum\n\n cons = ({'type': 'eq', 'fun': constraint})\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Local optimization with COBYLA (better for constrained optimization)\n result_slsqp = minimize(\n objective,\n best_h_de,\n method='COBYLA',\n bounds=bounds,\n constraints=cons,\n tol=1e-6,\n options={'maxiter': 200, 'disp': False}\n )\n\n best_h = result_slsqp.x\n best_c5 = evaluate_C5(best_h, dx)\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 306.81829041196033,
65 "env/all/time/policy/min": 156.0612998008728,
66 "env/all/time/policy/max": 409.4239983558655,
67 "env/all/time/env_step": 2136.9269451126456,
68 "env/all/time/env_step/min": 0.0037097930908203125,
69 "env/all/time/env_step/max": 4452.141984701157,
70 "env/all/time/reward_compute": 3.3993273973464966e-07,
71 "env/all/time/reward_compute/min": 2.0489096641540527e-07,
72 "env/all/time/reward_compute/max": 4.954636096954346e-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": 0.03543318808078766,
77 "advantage/min": -0.7645359039306641,
78 "advantage/max": 5.536941051483154,
79 "time/assemble_training_data": 9.172413349151611,
80 "time/kl_vs_base": 80.21165490150452,
81 "kl_policy_base": 0.0008720590267330408,
82 "time/train": 552.401043176651,
83 "time/save_checkpoint": 7.26540732383728,
84 "time/total": 5527.769442558289
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