Views
No views yet
erdos. Checkpoint saved
after training step 41 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 41,
3 "progress/batch": 41,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.84,
6 "puct/buffer_size": 662,
7 "puct/sampled_size": 8,
8 "puct/T": 20992,
9 "puct/scale_last": 0.1190591760561926,
10 "puct/buffer_value/mean": -0.3832448428228312,
11 "puct/buffer_value/std": 0.015341820002774913,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809408239438133,
14 "puct/buffer_timestep/mean": 19.70392749244713,
15 "puct/buffer_timestep/std": 11.97214603349419,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 40.0,
18 "puct/buffer_construction_len/mean": 79.31570996978851,
19 "puct/buffer_construction_len/std": 8.629064269994819,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.380940900660746,
23 "puct/sampled_value/std": 7.31690163313683e-08,
24 "puct/sampled_value/min": -0.38094099587026,
25 "puct/sampled_value/max": -0.38094084265840705,
26 "puct/sampled_timestep/mean": 40.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 40.0,
29 "puct/sampled_timestep/max": 40.0,
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": 3522.408773422241,
35 "env/all/ac_tokens_per_turn": 9387.462890625,
36 "env/all/ob_tokens_per_turn": 1252.625,
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": 4806381,
41 "env/all/total_ob_tokens": 641344,
42 "env/all/time/sampling_mean": 288.4399616513401,
43 "env/all/time/sampling_max": 403.1496572494507,
44 "env/all/time/env_step_mean": 1214.7009836812504,
45 "env/all/time/env_step_max": 3123.493305683136,
46 "env/all/reward/mean": 0.9777727766415181,
47 "env/all/reward/max": 2.625079472452571,
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": 0.9777727766415181,
53 "env/all/correctness": 0.375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.3841318950869635,
57 "env/all/raw_score/min": 0.380940837884394,
58 "env/all/raw_score/max": 0.5276867676323372,
59 "env/all/initial_raw_score": -0.38094090066074593,
60 "env/all/initial_raw_score/min": -0.38094099587026,
61 "env/all/initial_raw_score/max": -0.38094084265840705,
62 "env/all/msg": "Success; raw_score=0.38094161888793965",
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 # Consider all possible shifts\n max_corr = np.max(corr) * dx\n return max_corr\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 = 80\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n \n # Use provided initial_h_values if available\n if 'initial_h_values' in globals() and len(initial_h_values) == n_points:\n initial_guess = initial_h_values.copy()\n else:\n # Fallback to a random binary initial guess\n indices = np.random.choice(n_points, size=int(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 # Evaluate the max overlap\n max_c5 = evaluate_c5(h, dx)\n # Penalize constraint violation strongly\n sum_h = np.sum(h)\n penalty = 100 * (sum_h - required_sum) ** 2\n return max_c5 + penalty\n\n # Differential evolution with stronger parameters\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=50,\n maxiter=800,\n tol=1e-5,\n mutation=(0.3, 0.6),\n recombination=0.9,\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": 288.4399616513401,
65 "env/all/time/policy/min": 139.7986352443695,
66 "env/all/time/policy/max": 403.1496572494507,
67 "env/all/time/env_step": 1214.7009836812504,
68 "env/all/time/env_step/min": 0.0054929256439208984,
69 "env/all/time/env_step/max": 3123.493305683136,
70 "env/all/time/reward_compute": 3.5669654607772827e-07,
71 "env/all/time/reward_compute/min": 2.1606683731079102e-07,
72 "env/all/time/reward_compute/max": 5.62518835067749e-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": 0.04245075583457947,
77 "advantage/min": -1.0,
78 "advantage/max": 27.785701751708984,
79 "time/assemble_training_data": 7.764923572540283,
80 "time/kl_vs_base": 78.04014468193054,
81 "kl_policy_base": 0.0009140231413766742,
82 "time/train": 537.2747433185577,
83 "time/save_checkpoint": 19.683258295059204,
84 "time/total": 4166.660582065582
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