Views
No views yet
erdos. Checkpoint saved
after training step 20 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 20,
3 "progress/batch": 20,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.42,
6 "puct/buffer_size": 328,
7 "puct/sampled_size": 8,
8 "puct/T": 10240,
9 "puct/scale_last": 0.11905844420591044,
10 "puct/buffer_value/mean": -0.38525153659316436,
11 "puct/buffer_value/std": 0.020856672109041283,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38094155579409544,
14 "puct/buffer_timestep/mean": 9.24390243902439,
15 "puct/buffer_timestep/std": 5.921356501659915,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 19.0,
18 "puct/buffer_construction_len/mean": 78.6189024390244,
19 "puct/buffer_construction_len/std": 12.219706545725744,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.380941556290541,
23 "puct/sampled_value/std": 2.8629732001782723e-10,
24 "puct/sampled_value/min": -0.38094155659135764,
25 "puct/sampled_value/max": -0.38094155579409544,
26 "puct/sampled_timestep/mean": 19.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 19.0,
29 "puct/sampled_timestep/max": 19.0,
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": 2789.460270881653,
35 "env/all/ac_tokens_per_turn": 9539.275390625,
36 "env/all/ob_tokens_per_turn": 1459.375,
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": 4884109,
41 "env/all/total_ob_tokens": 747200,
42 "env/all/time/sampling_mean": 524.1634684223682,
43 "env/all/time/sampling_max": 758.0331926345825,
44 "env/all/time/env_step_mean": 459.96511635836214,
45 "env/all/time/env_step_max": 2045.6548254489899,
46 "env/all/reward/mean": 0.7643624283154304,
47 "env/all/reward/max": 2.6250745419877473,
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": 0.7643624283154304,
53 "env/all/correctness": 0.298828125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.3928831232553253,
57 "env/all/raw_score/min": 0.380941553374724,
58 "env/all/raw_score/max": 0.5925080303747119,
59 "env/all/initial_raw_score": -0.38094155629054105,
60 "env/all/initial_raw_score/min": -0.38094155659135764,
61 "env/all/initial_raw_score/max": -0.38094155579409544,
62 "env/all/msg": "Success; raw_score=0.38094289664684633",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\nfrom scipy.signal import correlate\n\ndef evaluate_C5(h, dx):\n h1 = 1.0 - h\n corr = correlate(h, h1, mode='full')\n max_corr_value = np.max(corr) * dx\n return max_corr_value\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n import random\n random.seed(seed)\n \n # Use the initial_h_values as the canonical starting point\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2\n best_h = initial_h_values.copy()\n best_c5 = float('inf')\n\n def objective(x):\n corr = correlate(x, 1.0 - x, mode='full')\n max_corr = np.max(corr) * dx\n return max_corr\n\n def constraint(x):\n return np.sum(x) - required_sum\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Optimization with COBYLA method\n def optimize(initial_guess):\n result = minimize(\n fun=objective,\n x0=initial_guess,\n method='COBYLA',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint}],\n options={'ftol': 1e-8, 'maxiter': 10000, 'disp': False}\n )\n return result.x, result.fun\n\n # Evaluate the initial guess\n h_initial, c5_initial = optimize(best_h)\n best_c5 = c5_initial\n best_h = h_initial\n\n # Try a few randomized starting points\n num_initial_guesses = 3\n for _ in range(num_initial_guesses):\n # Generate a random guess that satisfies the integral constraint\n h_rand = np.random.rand(n_points)\n h_rand_sum = np.sum(h_rand)\n h_rand *= required_sum / h_rand_sum\n h_rand, c5_rand = optimize(h_rand)\n if c5_rand < best_c5:\n best_c5 = c5_rand\n best_h = h_rand\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 524.1634684223682,
65 "env/all/time/policy/min": 225.63932013511658,
66 "env/all/time/policy/max": 758.0331926345825,
67 "env/all/time/env_step": 459.96511635836214,
68 "env/all/time/env_step/min": 0.00505828857421875,
69 "env/all/time/env_step/max": 2045.6548254489899,
70 "env/all/time/reward_compute": 3.548339009284973e-07,
71 "env/all/time/reward_compute/min": 1.9371509552001953e-07,
72 "env/all/time/reward_compute/max": 5.066394805908203e-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": 0.03019561618566513,
77 "advantage/min": -0.9841427803039551,
78 "advantage/max": 8.202388763427734,
79 "time/assemble_training_data": 8.923372030258179,
80 "time/kl_vs_base": 132.8651421070099,
81 "kl_policy_base": 0.0008594145765528083,
82 "time/train": 1029.4861009120941,
83 "time/save_checkpoint": 8.087226390838623,
84 "time/total": 3972.58469247818
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