Views
No views yet
erdos. Checkpoint saved
after training step 49 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 49,
3 "progress/batch": 49,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 1.0,
6 "puct/buffer_size": 784,
7 "puct/sampled_size": 8,
8 "puct/T": 25088,
9 "puct/scale_last": 0.12405074150219608,
10 "puct/buffer_value/mean": -0.3832127518222866,
11 "puct/buffer_value/std": 0.015133239842634054,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38094051871019485,
14 "puct/buffer_timestep/mean": 23.556122448979593,
15 "puct/buffer_timestep/std": 14.224221698348973,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 48.0,
18 "puct/buffer_construction_len/mean": 79.54974489795919,
19 "puct/buffer_construction_len/std": 8.469939337675479,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 160.0,
22 "puct/sampled_value/mean": -0.38094064579488285,
23 "puct/sampled_value/std": 6.679176239074958e-08,
24 "puct/sampled_value/min": -0.38094073193798067,
25 "puct/sampled_value/max": -0.38094051871019485,
26 "puct/sampled_timestep/mean": 48.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 48.0,
29 "puct/sampled_timestep/max": 48.0,
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": 2885.9708108901978,
35 "env/all/ac_tokens_per_turn": 9247.423828125,
36 "env/all/ob_tokens_per_turn": 1284.375,
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": 4734681,
41 "env/all/total_ob_tokens": 657600,
42 "env/all/time/sampling_mean": 283.78204646054655,
43 "env/all/time/sampling_max": 386.86468505859375,
44 "env/all/time/env_step_mean": 900.9445394580252,
45 "env/all/time/env_step_max": 2500.476953983307,
46 "env/all/reward/mean": 0.4387160390476521,
47 "env/all/reward/max": 2.6250816200426703,
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": 0.4387160390476521,
53 "env/all/correctness": 0.16796875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.3833170894615982,
57 "env/all/raw_score/min": 0.3809405262351153,
58 "env/all/raw_score/max": 0.5108707817409756,
59 "env/all/initial_raw_score": -0.3809406457948829,
60 "env/all/initial_raw_score/min": -0.38094073193798067,
61 "env/all/initial_raw_score/max": -0.38094051871019485,
62 "env/all/msg": "C5 mismatch: reported 0.01537333, computed 0.96895448",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n overlap_start = n_points - 1\n overlap_end = 2 * n_points - 1\n valid_corr = corr[overlap_start:overlap_end]\n max_corr = np.max(valid_corr)\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n\n # Use provided initial guess if available\n if 'initial_h_values' in globals() and len(initial_h_values) > 0:\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n initial_guess = initial_h_values.copy()\n else:\n # Fallback: 80 points with random 0/1 values for more diverse initial search\n n_points = 80\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n indices = np.random.choice(n_points, size=n_points // 2, replace=False)\n initial_guess = np.zeros(n_points)\n initial_guess[indices] = 1.0\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(x):\n h = np.array(x)\n sum_h = np.sum(h)\n # Penalize sum deviation with higher weight for strict constraint\n penalty = 1e3 * (sum_h - required_sum) ** 2\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Run differential evolution for global exploration with more aggressive parameters\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin', # More robust for exploration\n popsize=100, # Larger population for better exploration\n maxiter=400, # Increased iterations for better exploration\n tol=1e-4,\n mutation=(0.8, 1.0), # Broad mutation for exploration\n recombination=0.95, # Higher recombination for mixing\n seed=seed,\n x0=initial_guess,\n polish=True # Refine the best individual after evolution\n )\n\n best_h = result_de.x\n best_c5 = evaluate_c5(best_h, dx)\n\n # Local refinement with SLSQP and explicit equality constraint\n constraints = [\n {'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum}\n ]\n result_slsqp = minimize(\n objective,\n best_h,\n bounds=bounds,\n method='SLSQP',\n constraints=constraints,\n tol=1e-8,\n options={'maxiter': 700, 'ftol': 1e-7, 'eps': 1e-7}\n )\n\n final_h = result_slsqp.x\n final_c5 = evaluate_c5(final_h, dx)\n\n return final_h, final_c5, n_points\n```",
64 "env/all/time/policy": 283.78204646054655,
65 "env/all/time/policy/min": 149.9666233062744,
66 "env/all/time/policy/max": 386.86468505859375,
67 "env/all/time/env_step": 900.9445394580252,
68 "env/all/time/env_step/min": 0.004969358444213867,
69 "env/all/time/env_step/max": 2500.476953983307,
70 "env/all/time/reward_compute": 2.5704503059387207e-07,
71 "env/all/time/reward_compute/min": 1.7136335372924805e-07,
72 "env/all/time/reward_compute/max": 3.762543201446533e-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": 0.04158974066376686,
77 "advantage/min": -0.874153196811676,
78 "advantage/max": 9.887174606323242,
79 "time/assemble_training_data": 8.933504104614258,
80 "time/kl_vs_base": 78.3884949684143,
81 "kl_policy_base": 0.0009776601800695062,
82 "time/train": 529.8142125606537,
83 "time/save_checkpoint": 16.5112144947052,
84 "time/total": 3520.912221431732
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