Views
No views yet
erdos. Checkpoint saved
after training step 36 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 36,
3 "progress/batch": 36,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.74,
6 "puct/buffer_size": 583,
7 "puct/sampled_size": 8,
8 "puct/T": 18432,
9 "puct/scale_last": 0.1190591760561926,
10 "puct/buffer_value/mean": -0.38338097782075253,
11 "puct/buffer_value/std": 0.01578854754892028,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809408239438133,
14 "puct/buffer_timestep/mean": 17.224699828473412,
15 "puct/buffer_timestep/std": 10.534397168761531,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 35.0,
18 "puct/buffer_construction_len/mean": 79.22298456260721,
19 "puct/buffer_construction_len/std": 9.191222974753213,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38094148966339847,
23 "puct/sampled_value/std": 2.0944109502360107e-08,
24 "puct/sampled_value/min": -0.38094150048859254,
25 "puct/sampled_value/max": -0.38094143664669744,
26 "puct/sampled_timestep/mean": 35.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 35.0,
29 "puct/sampled_timestep/max": 35.0,
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": 4041.7924168109894,
35 "env/all/ac_tokens_per_turn": 9737.888671875,
36 "env/all/ob_tokens_per_turn": 1232.125,
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": 4985799,
41 "env/all/total_ob_tokens": 630848,
42 "env/all/time/sampling_mean": 304.997365728952,
43 "env/all/time/sampling_max": 418.7341158390045,
44 "env/all/time/env_step_mean": 1581.0190776800737,
45 "env/all/time/env_step_max": 3612.9030978679657,
46 "env/all/reward/mean": 0.9377604023718199,
47 "env/all/reward/max": 2.625079310796865,
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": 0.9377604023718199,
53 "env/all/correctness": 0.359375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.38366834323197174,
57 "env/all/raw_score/min": 0.3809408613432135,
58 "env/all/raw_score/max": 0.483723984144869,
59 "env/all/initial_raw_score": -0.38094148966339847,
60 "env/all/initial_raw_score/min": -0.38094150048859254,
61 "env/all/initial_raw_score/max": -0.38094143664669744,
62 "env/all/msg": "C5 mismatch: reported 0.00000000, computed 0.72222222",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\nfrom math import inf\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 valid_corr = corr[-n_points:] # Valid shifts from m=0 to m=n_points-1\n max_c5 = np.max(valid_corr) * dx\n return max_c5\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n import random\n random.seed(seed)\n\n # Generate structured initial guess\n m = 5 # Number of intervals\n n_points = 80\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n # Generate intervals\n s = 1.0 / (m - 1) # spacing between intervals\n a = 1.0 / m # length of each interval\n start_positions = [i * s for i in range(m)]\n h_values = np.zeros(n_points)\n for i in range(n_points):\n x = i * dx\n for j in range(m):\n if start_positions[j] <= x < start_positions[j] + a:\n h_values[i] = 1.0\n break\n\n # Use the structured guess as initial guess\n initial_guess = h_values.copy()\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 c5_val = evaluate_c5(h, dx)\n # Penalties are not used due to constraint satisfaction by initial guess\n return c5_val\n\n # Run differential evolution with optimized parameters\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=200, # Increased population for better diversity\n maxiter=300, # Increased max iterations for better exploration\n tol=1e-9,\n mutation=(0.5, 0.9), # Wider mutation range\n recombination=0.95, # Higher recombination to improve convergence\n seed=seed,\n x0=initial_guess\n )\n\n best_h = result.x\n best_c5 = evaluate_c5(best_h, dx)\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 304.997365728952,
65 "env/all/time/policy/min": 137.52771925926208,
66 "env/all/time/policy/max": 418.7341158390045,
67 "env/all/time/env_step": 1581.0190776800737,
68 "env/all/time/env_step/min": 0.005166292190551758,
69 "env/all/time/env_step/max": 3612.9030978679657,
70 "env/all/time/reward_compute": 3.0314549803733826e-07,
71 "env/all/time/reward_compute/min": 2.0489096641540527e-07,
72 "env/all/time/reward_compute/max": 3.7997961044311523e-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": 0.02714661881327629,
77 "advantage/min": -1.0,
78 "advantage/max": 12.02221965789795,
79 "time/assemble_training_data": 6.839499235153198,
80 "time/kl_vs_base": 84.22841262817383,
81 "kl_policy_base": 0.0008681275648996234,
82 "time/train": 556.4792191982269,
83 "time/save_checkpoint": 16.715221881866455,
84 "time/total": 4707.246897220612
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