Views
No views yet
erdos. Checkpoint saved
after training step 7 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 7,
3 "progress/batch": 7,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.16,
6 "puct/buffer_size": 120,
7 "puct/sampled_size": 8,
8 "puct/T": 3584,
9 "puct/scale_last": 0.030273321240342432,
10 "puct/buffer_value/mean": -0.39083983980935305,
11 "puct/buffer_value/std": 0.029620012533741216,
12 "puct/buffer_value/min": -0.5130522804051018,
13 "puct/buffer_value/max": -0.3811063325170931,
14 "puct/buffer_timestep/mean": 2.7333333333333334,
15 "puct/buffer_timestep/std": 2.174600857373345,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 6.0,
18 "puct/buffer_construction_len/mean": 60.71666666666667,
19 "puct/buffer_construction_len/std": 20.95359290326018,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 100.0,
22 "puct/sampled_value/mean": -0.38122511123128666,
23 "puct/sampled_value/std": 6.377289348872367e-05,
24 "puct/sampled_value/min": -0.38129657028434244,
25 "puct/sampled_value/max": -0.3811063325170931,
26 "puct/sampled_timestep/mean": 6.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 6.0,
29 "puct/sampled_timestep/max": 6.0,
30 "puct/sampled_construction_len/mean": 67.0,
31 "puct/sampled_construction_len/std": 16.658331248957683,
32 "puct/sampled_construction_len/min": 53.0,
33 "puct/sampled_construction_len/max": 100.0,
34 "time/sampling": 1881.7153153419495,
35 "env/all/ac_tokens_per_turn": 8522.38671875,
36 "env/all/ob_tokens_per_turn": 1585.125,
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": 4363462,
41 "env/all/total_ob_tokens": 811584,
42 "env/all/time/sampling_mean": 458.4608020358719,
43 "env/all/time/sampling_max": 646.5195879936218,
44 "env/all/time/env_step_mean": 133.166764610447,
45 "env/all/time/env_step_max": 1231.0708463191986,
46 "env/all/reward/mean": 1.0172504827547817,
47 "env/all/reward/max": 2.624553272365123,
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": 1.0172504827547817,
53 "env/all/correctness": 0.392578125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.38691504732724175,
57 "env/all/raw_score/min": 0.38101721320875104,
58 "env/all/raw_score/max": 0.5000000000000088,
59 "env/all/initial_raw_score": -0.3812251112312867,
60 "env/all/initial_raw_score/min": -0.38129657028434244,
61 "env/all/initial_raw_score/max": -0.3811063325170931,
62 "env/all/msg": "Success; raw_score=0.38108646162479093",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import basinhopping, minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 100\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Start from the better known initial guess\n h_initial = np.array(initial_h_values, dtype=np.float64)\n h_initial[-1] = n_points / 2 - np.sum(h_initial[:-1]) # Ensures the sum constraint\n\n def objective(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n max_corr = np.max(corr)\n return max_corr * dx\n\n def constraint_func(h):\n return np.sum(h) - n_points / 2\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Basin-hopping for global exploration\n result_bh = basinhopping(\n objective,\n h_initial,\n niter=150, # More iterations for global exploration\n T=1.5, # Moderately high temperature to allow more exploration\n stepsize=0.2,\n minimizer_kwargs={\n 'method': 'SLSQP',\n 'bounds': bounds,\n 'constraints': [{'type': 'eq', 'fun': constraint_func}],\n 'options': {'ftol': 1e-10, 'maxiter': 500, 'disp': False}\n },\n seed=seed\n )\n\n best_h = result_bh.x\n best_c5 = result_bh.fun\n\n # Refine with SLSQP\n constraints = [{'type': 'eq', 'fun': constraint_func}]\n res_slsqp = minimize(\n fun=objective,\n x0=best_h,\n method='SLSQP',\n bounds=bounds,\n constraints=constraints,\n options={'ftol': 1e-10, 'maxiter': 500}\n )\n\n best_h, best_c5 = res_slsqp.x, res_slsqp.fun\n\n # Evaluate all cross-correlation values to find the lags with the largest overlaps\n def get_max_lag(h_array):\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n max_idx = np.argmax(corr)\n return max_idx, corr[max_idx] * dx\n\n max_idx, max_overlap_val = get_max_lag(best_h)\n\n # Apply targeted perturbations to the lag with the highest overlap\n for _ in range(20):\n h_perturbed = best_h + np.random.uniform(-0.05, 0.05, size=n_points)\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n h_sum = np.sum(h_perturbed[:-1])\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - h_sum))\n res_perturbed = minimize(\n fun=objective,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=constraints,\n options={'ftol': 1e-10, 'maxiter': 200}\n )\n if res_perturbed.fun < best_c5:\n best_h, best_c5 = res_perturbed.x, res_perturbed.fun\n\n # Final fine-tuning\n for _ in range(20):\n h_perturbed = best_h + np.random.uniform(-0.01, 0.01, size=n_points)\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n h_sum = np.sum(h_perturbed[:-1])\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - h_sum))\n res_perturbed = minimize(\n fun=objective,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=constraints,\n options={'ftol': 1e-10, 'maxiter': 100}\n )\n if res_perturbed.fun < best_c5:\n best_h, best_c5 = res_perturbed.x, res_perturbed.fun\n\n return (best_h, best_c5, n_points)\n```",
64 "env/all/time/policy": 458.4608020358719,
65 "env/all/time/policy/min": 207.1914520263672,
66 "env/all/time/policy/max": 646.5195879936218,
67 "env/all/time/env_step": 133.166764610447,
68 "env/all/time/env_step/min": 0.0052721500396728516,
69 "env/all/time/env_step/max": 1231.0708463191986,
70 "env/all/time/reward_compute": 4.931353032588959e-07,
71 "env/all/time/reward_compute/min": 1.5273690223693848e-07,
72 "env/all/time/reward_compute/max": 1.6614794731140137e-06,
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": 0.023009292781352997,
77 "advantage/min": -1.0,
78 "advantage/max": 3.4012036323547363,
79 "time/assemble_training_data": 9.01777696609497,
80 "time/kl_vs_base": 125.17944860458374,
81 "kl_policy_base": 0.0006696797790937126,
82 "time/train": 926.7443289756775,
83 "time/save_checkpoint": 9.232323169708252,
84 "time/total": 2953.0930066108704
85}[2026-07-09T06:24:18+00:00] job=1812626 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:02:36+00:00] job=1813125 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:32:13+00:00] job=1813126 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T10:04:31+00:00] job=1813623 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=yes