Views
No views yet
erdos. Checkpoint saved
after training step 32 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 32,
3 "progress/batch": 32,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.66,
6 "puct/buffer_size": 519,
7 "puct/sampled_size": 8,
8 "puct/T": 16384,
9 "puct/scale_last": 0.1190591760561926,
10 "puct/buffer_value/mean": -0.3836817968969181,
11 "puct/buffer_value/std": 0.016709081947286143,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809408239438133,
14 "puct/buffer_timestep/mean": 15.217726396917149,
15 "puct/buffer_timestep/std": 9.370802532873716,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 31.0,
18 "puct/buffer_construction_len/mean": 79.1271676300578,
19 "puct/buffer_construction_len/std": 9.73716307614876,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.3809415179015716,
23 "puct/sampled_value/std": 1.0322199560808305e-08,
24 "puct/sampled_value/min": -0.38094152612665116,
25 "puct/sampled_value/max": -0.38094149651651277,
26 "puct/sampled_timestep/mean": 31.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 31.0,
29 "puct/sampled_timestep/max": 31.0,
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": 4038.2425808906555,
35 "env/all/ac_tokens_per_turn": 9243.583984375,
36 "env/all/ob_tokens_per_turn": 1242.625,
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": 4732715,
41 "env/all/total_ob_tokens": 636224,
42 "env/all/time/sampling_mean": 285.9204009906389,
43 "env/all/time/sampling_max": 379.57638597488403,
44 "env/all/time/env_step_mean": 1645.1548493718728,
45 "env/all/time/env_step_max": 3657.911336660385,
46 "env/all/reward/mean": 0.48197417668720716,
47 "env/all/reward/max": 2.6250752452047195,
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": 0.48197417668720716,
53 "env/all/correctness": 0.185546875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.38592307533305015,
57 "env/all/raw_score/min": 0.38094145132638335,
58 "env/all/raw_score/max": 0.5,
59 "env/all/initial_raw_score": -0.3809415179015716,
60 "env/all/initial_raw_score/min": -0.38094152612665116,
61 "env/all/initial_raw_score/max": -0.38094149651651277,
62 "env/all/msg": "Success; raw_score=0.3809415738722977",
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 one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n max_c5 = np.max(corr) * dx # consider all shifts\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 n_points = len(initial_h_values) if initial_h_values is not None else 80\n dx = 2.0 / n_points\n required_sum = n_points / 2\n\n # Initial guess: use provided initial_h_values if available\n if initial_h_values is not None:\n initial_guess = initial_h_values.copy()\n else:\n # Fallback to uniform random binary initial guess\n indices = np.random.choice(n_points, size=int(required_sum), 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 penalty = 1000.0 * (sum_h - required_sum) ** 2\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Use best1bin strategy for better convergence\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin', # balance between exploration and exploitation\n popsize=80, # increased population size\n maxiter=1000, # more iterations for thorough search\n tol=1e-6,\n mutation=(0.2, 0.4),\n recombination=0.7, # increased recombination for diversity\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": 285.9204009906389,
65 "env/all/time/policy/min": 142.18357157707214,
66 "env/all/time/policy/max": 379.57638597488403,
67 "env/all/time/env_step": 1645.1548493718728,
68 "env/all/time/env_step/min": 0.004938840866088867,
69 "env/all/time/env_step/max": 3657.911336660385,
70 "env/all/time/reward_compute": 3.0407682061195374e-07,
71 "env/all/time/reward_compute/min": 2.2724270820617676e-07,
72 "env/all/time/reward_compute/max": 4.917383193969727e-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": 0.03465929627418518,
77 "advantage/min": -0.7505736351013184,
78 "advantage/max": 7.753525733947754,
79 "time/assemble_training_data": 7.78698468208313,
80 "time/kl_vs_base": 77.3703064918518,
81 "kl_policy_base": 0.0008580085122957826,
82 "time/train": 527.1687133312225,
83 "time/save_checkpoint": 14.111188888549805,
84 "time/total": 4666.363662242889
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