Views
No views yet
erdos. Checkpoint saved
after training step 3 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 3,
3 "progress/batch": 3,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.08,
6 "puct/buffer_size": 56,
7 "puct/sampled_size": 8,
8 "puct/T": 1536,
9 "puct/scale_last": 0.023671422269292475,
10 "puct/buffer_value/mean": -0.40182358878675295,
11 "puct/buffer_value/std": 0.043281374561392645,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38145954985425273,
14 "puct/buffer_timestep/mean": 0.7142857142857143,
15 "puct/buffer_timestep/std": 1.0301575072754257,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 2.0,
18 "puct/buffer_construction_len/mean": 76.125,
19 "puct/buffer_construction_len/std": 27.050919247649553,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38170830476968076,
23 "puct/sampled_value/std": 0.00011903194570160251,
24 "puct/sampled_value/min": -0.3818563605361555,
25 "puct/sampled_value/max": -0.38145954985425273,
26 "puct/sampled_timestep/mean": 2.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 2.0,
29 "puct/sampled_timestep/max": 2.0,
30 "puct/sampled_construction_len/mean": 71.75,
31 "puct/sampled_construction_len/std": 12.346558224865746,
32 "puct/sampled_construction_len/min": 56.0,
33 "puct/sampled_construction_len/max": 86.0,
34 "time/sampling": 1784.6060194969177,
35 "env/all/ac_tokens_per_turn": 9249.5078125,
36 "env/all/ob_tokens_per_turn": 1262.0,
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": 4735748,
41 "env/all/total_ob_tokens": 646144,
42 "env/all/time/sampling_mean": 493.3677526395768,
43 "env/all/time/sampling_max": 712.885725736618,
44 "env/all/time/env_step_mean": 53.228829064872116,
45 "env/all/time/env_step_max": 1100.1254024505615,
46 "env/all/reward/mean": 0.8192612580261633,
47 "env/all/reward/max": 2.623154128705225,
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": 0.8192612580261633,
53 "env/all/correctness": 0.32421875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.39900099357774194,
57 "env/all/raw_score/min": 0.38122044100475844,
58 "env/all/raw_score/max": 0.5183051091904262,
59 "env/all/initial_raw_score": -0.38170830476968076,
60 "env/all/initial_raw_score/min": -0.3818563605361555,
61 "env/all/initial_raw_score/max": -0.38145954985425273,
62 "env/all/msg": "Success; raw_score=0.38145748334157176",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import basinhopping\n\ndef run(seed=42, budget_s=1000, **kwargs):\n # Use the provided initial construction\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n h_initial = initial_h_values.astype(float)\n\n # Objective function: minimize the max overlap integral\n def objective(h):\n h = np.array(h)\n h1 = 1.0 - h\n corr = np.correlate(h, h1, mode='full')\n return np.max(corr) * dx\n\n # Constraints: integral of h(x) over [0,2] must be 1\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n constraints = {'type': 'eq', 'fun': lambda h: np.sum(h) - n_points / 2}\n\n # Run basinhopping with enhanced parameters\n result = basinhopping(\n objective,\n x0=h_initial,\n niter=300, # Increase number of iterations\n T=0.2, # Lower temperature for fine convergence\n stepsize=0.5, # Larger step size for exploration\n minimizer_kwargs={\n 'method': 'SLSQP',\n 'bounds': bounds,\n 'constraints': constraints,\n 'options': {\n 'ftol': 1e-8, # Tight tolerance\n 'maxiter': 300 # Allow more iterations per local solver\n }\n }\n )\n\n h_values = result.x\n c5_bound = result.fun\n return h_values, c5_bound, n_points\n```",
64 "env/all/time/policy": 493.3677526395768,
65 "env/all/time/policy/min": 170.54231095314026,
66 "env/all/time/policy/max": 712.885725736618,
67 "env/all/time/env_step": 53.228829064872116,
68 "env/all/time/env_step/min": 0.00589299201965332,
69 "env/all/time/env_step/max": 1100.1254024505615,
70 "env/all/time/reward_compute": 3.976747393608093e-07,
71 "env/all/time/reward_compute/min": 3.0547380447387695e-07,
72 "env/all/time/reward_compute/max": 5.774199962615967e-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": 0.025713184848427773,
77 "advantage/min": -0.9931661486625671,
78 "advantage/max": 3.8230905532836914,
79 "time/assemble_training_data": 9.234843730926514,
80 "time/kl_vs_base": 126.4726037979126,
81 "kl_policy_base": 0.0006338226376101375,
82 "time/train": 979.1366686820984,
83 "time/save_checkpoint": 13.365399837493896,
84 "time/total": 2914.039454936981
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