Views
No views yet
erdos. Checkpoint saved
after training step 17 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 17,
3 "progress/batch": 17,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.36,
6 "puct/buffer_size": 280,
7 "puct/sampled_size": 8,
8 "puct/T": 8704,
9 "puct/scale_last": 0.11907305060535461,
10 "puct/buffer_value/mean": -0.38612923746671135,
11 "puct/buffer_value/std": 0.022193130251737005,
12 "puct/buffer_value/min": -0.5130522804051018,
13 "puct/buffer_value/max": -0.3809269493946454,
14 "puct/buffer_timestep/mean": 7.742857142857143,
15 "puct/buffer_timestep/std": 5.055932058443017,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 16.0,
18 "puct/buffer_construction_len/mean": 77.7,
19 "puct/buffer_construction_len/std": 25.73371884734668,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 200.0,
22 "puct/sampled_value/mean": -0.3809293326321679,
23 "puct/sampled_value/std": 2.0253461349849064e-06,
24 "puct/sampled_value/min": -0.38093150577749546,
25 "puct/sampled_value/max": -0.3809269493946454,
26 "puct/sampled_timestep/mean": 16.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 16.0,
29 "puct/sampled_timestep/max": 16.0,
30 "puct/sampled_construction_len/mean": 100.0,
31 "puct/sampled_construction_len/std": 0.0,
32 "puct/sampled_construction_len/min": 100.0,
33 "puct/sampled_construction_len/max": 100.0,
34 "time/sampling": 2913.670128583908,
35 "env/all/ac_tokens_per_turn": 9341.662109375,
36 "env/all/ob_tokens_per_turn": 1731.75,
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": 4782931,
41 "env/all/total_ob_tokens": 886656,
42 "env/all/time/sampling_mean": 529.451841423288,
43 "env/all/time/sampling_max": 757.0236597061157,
44 "env/all/time/env_step_mean": 767.5798722803593,
45 "env/all/time/env_step_max": 2172.275707244873,
46 "env/all/reward/mean": 0.8416086726330138,
47 "env/all/reward/max": 2.625183488590748,
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": 0.8416086726330138,
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": 0.38719497809409564,
57 "env/all/raw_score/min": 0.380925744083887,
58 "env/all/raw_score/max": 0.7832916383362131,
59 "env/all/initial_raw_score": -0.380929332632168,
60 "env/all/initial_raw_score/min": -0.38093150577749546,
61 "env/all/initial_raw_score/max": -0.3809269493946454,
62 "env/all/msg": "C5 mismatch: reported 0.02000000, computed 0.49843686",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 100\n dx = 2.0 / n_points\n target_sum = n_points / 2\n\n # Define the objective function for DE\n def objective(vars):\n h = np.zeros(n_points)\n h[:-1] = vars\n h[-1] = np.clip(target_sum - np.sum(vars), 0.0, 1.0)\n corr = np.correlate(h, 1 - h, mode='full')\n return np.max(corr) * dx\n\n # Bounds for the first n_points - 1 variables\n bounds = [(0.0, 1.0) for _ in range(n_points - 1)]\n\n # Run differential evolution for global optimization\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=20,\n maxiter=500,\n tol=1e-10,\n mutation=(0.5, 1),\n recombination=0.9,\n seed=seed\n )\n\n best_h = np.zeros(n_points)\n best_h[:-1] = result_de.x\n best_h[-1] = np.clip(target_sum - np.sum(best_h[:-1]), 0.0, 1.0)\n best_c5 = result_de.fun\n\n # Local refinement using targeted perturbation and SLSQP\n for _ in range(400):\n h_array = best_h.copy()\n corr = np.correlate(h_array, 1 - h_array, mode='full')\n max_k_index = np.argmax(corr)\n shift = max_k_index - n_points + 1\n\n # Compute contributions to the overlap\n contributions = []\n for i in range(n_points):\n x = i * dx\n shifted_x = x + shift\n if 0 <= shifted_x <= 2:\n j = int(round(shifted_x / dx))\n if 0 <= j < n_points:\n contrib = h_array[i] * (1 - h_array[j])\n else:\n contrib = 0\n else:\n contrib = 0\n contributions.append(contrib)\n\n # Perturb the top contributors\n top_indices = np.argsort(contributions)[::-1][:20]\n perturbation = np.random.uniform(-0.03, 0.03, size=n_points)\n for i in top_indices:\n h_array[i] += perturbation[i]\n h_array[i] = np.clip(h_array[i], 0.0, 1.0)\n h_array[-1] = np.clip(target_sum - np.sum(h_array[:-1]), 0.0, 1.0)\n\n # Local minimization with SLSQP\n res = minimize(\n fun=objective,\n x0=h_array[:-1],\n bounds=[(0.0, 1.0)] * (n_points - 1),\n method='SLSQP',\n constraints=[{'type': 'eq', 'fun': lambda x: target_sum - np.sum(x)}],\n options={'ftol': 1e-10, 'maxiter': 200}\n )\n\n if res.fun < best_c5:\n best_h[:-1] = res.x\n best_h[-1] = np.clip(target_sum - np.sum(res.x), 0.0, 1.0)\n best_c5 = res.fun\n\n return (best_h, best_c5, n_points)\n```",
64 "env/all/time/policy": 529.451841423288,
65 "env/all/time/policy/min": 175.5647838115692,
66 "env/all/time/policy/max": 757.0236597061157,
67 "env/all/time/env_step": 767.5798722803593,
68 "env/all/time/env_step/min": 0.00672459602355957,
69 "env/all/time/env_step/max": 2172.275707244873,
70 "env/all/time/reward_compute": 3.259629011154175e-07,
71 "env/all/time/reward_compute/min": 2.2351741790771484e-07,
72 "env/all/time/reward_compute/max": 4.76837158203125e-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": 0.025077100843191147,
77 "advantage/min": -1.0,
78 "advantage/max": 3.345200538635254,
79 "time/assemble_training_data": 8.49018383026123,
80 "time/kl_vs_base": 130.37992787361145,
81 "kl_policy_base": 0.0008462729747407138,
82 "time/train": 1032.0425922870636,
83 "time/save_checkpoint": 9.041590213775635,
84 "time/total": 4095.1571803092957
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