Views
No views yet
erdos. Checkpoint saved
after training step 26 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 26,
3 "progress/batch": 26,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.54,
6 "puct/buffer_size": 424,
7 "puct/sampled_size": 8,
8 "puct/T": 13312,
9 "puct/scale_last": 0.11907473987989126,
10 "puct/buffer_value/mean": -0.38436514734900074,
11 "puct/buffer_value/std": 0.018201943520077275,
12 "puct/buffer_value/min": -0.5130522804051018,
13 "puct/buffer_value/max": -0.38092526012010874,
14 "puct/buffer_timestep/mean": 12.245283018867925,
15 "puct/buffer_timestep/std": 7.652612478761028,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 25.0,
18 "puct/buffer_construction_len/mean": 87.39622641509433,
19 "puct/buffer_construction_len/std": 28.62037776647796,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 200.0,
22 "puct/sampled_value/mean": -0.38092527428301504,
23 "puct/sampled_value/std": 6.875869323168489e-09,
24 "puct/sampled_value/min": -0.38092528188974767,
25 "puct/sampled_value/max": -0.38092526012010874,
26 "puct/sampled_timestep/mean": 25.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 25.0,
29 "puct/sampled_timestep/max": 25.0,
30 "puct/sampled_construction_len/mean": 137.5,
31 "puct/sampled_construction_len/std": 48.41229182759271,
32 "puct/sampled_construction_len/min": 100.0,
33 "puct/sampled_construction_len/max": 200.0,
34 "time/sampling": 3285.6169095039368,
35 "env/all/ac_tokens_per_turn": 9140.373046875,
36 "env/all/ob_tokens_per_turn": 1503.0,
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": 4679871,
41 "env/all/total_ob_tokens": 769536,
42 "env/all/time/sampling_mean": 493.082740210928,
43 "env/all/time/sampling_max": 771.7152693271637,
44 "env/all/time/env_step_mean": 978.2773189232685,
45 "env/all/time/env_step_max": 2562.9273443222046,
46 "env/all/reward/mean": 0.6005337988549904,
47 "env/all/reward/max": 2.6251867856801745,
48 "env/all/reward/min": 0.0,
49 "env/all/format": 0.998046875,
50 "env/all/format/min": 0.0,
51 "env/all/format/max": 1.0,
52 "env/all/reward": 0.6005337988549904,
53 "env/all/correctness": 0.244140625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.4133111509122273,
57 "env/all/raw_score/min": 0.38092526566220564,
58 "env/all/raw_score/max": 0.8585218545495176,
59 "env/all/initial_raw_score": -0.38092527428301504,
60 "env/all/initial_raw_score/min": -0.38092528188974767,
61 "env/all/initial_raw_score/max": -0.38092526012010874,
62 "env/all/msg": "Success; raw_score=0.4088676970095786",
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 = 400 # Increased for better resolution\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n desired_sum = n_points / 2.0\n\n # Handle initial guess\n try:\n if len(initial_h_values) == n_points:\n initial_h = initial_h_values.copy()\n else:\n # Random initial guess with exactly n_points/2 ones and zeros\n if n_points % 2 != 0:\n raise ValueError(\"n_points must be even for this initial guess\")\n h = np.random.choice([0, 1], size=n_points, p=[0.5, 0.5])\n # Adjust to ensure the sum constraint is met\n if np.sum(h) > desired_sum:\n h[-1] = 0\n elif np.sum(h) < desired_sum:\n h[-1] = 1\n initial_h = h\n except NameError:\n # Fallback to random initial guess\n initial_h = np.random.choice([0, 1], size=n_points, p=[0.5, 0.5])\n # Ensure sum constraint\n if np.sum(initial_h) > desired_sum:\n initial_h[-1] = 0\n elif np.sum(initial_h) < desired_sum:\n initial_h[-1] = 1\n\n # Objectives and constraints\n def compute_correlation_fft(h_array):\n n = len(h_array)\n reversed_1_minus_h = (1 - h_array)[::-1]\n h_fft = np.fft.fft(h_array, 2 * n - 1)\n reversed_fft = np.fft.fft(reversed_1_minus_h, 2 * n - 1)\n corr_fft = h_fft * reversed_fft\n corr = np.fft.ifft(corr_fft).real\n return corr\n\n def objective_with_penalty(h_array):\n corr = compute_correlation_fft(h_array)\n obj = np.max(corr) * dx\n sum_penalty = 1e4 * (np.sum(h_array) - desired_sum) ** 2\n return obj + sum_penalty\n\n def objective(h_array):\n corr = compute_correlation_fft(h_array)\n return np.max(corr) * dx\n\n # Bounds for all points to be in [0,1]\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Global optimization: Differential Evolution\n result_de = differential_evolution(\n objective_with_penalty,\n bounds,\n strategy='best1bin',\n popsize=20,\n maxiter=100,\n tol=1e-8,\n mutation=(0.5, 1),\n recombination=0.8,\n disp=False\n )\n\n best_h_de = result_de.x\n best_c5_de = objective_with_penalty(best_h_de)\n\n # Local refinement with SLSQP\n result_slsqp = minimize(\n fun=objective,\n x0=best_h_de,\n method='SLSQP',\n bounds=bounds,\n constraints=[{\n 'type': 'eq',\n 'fun': lambda h: np.sum(h) - desired_sum\n }],\n options={\n 'ftol': 1e-8,\n 'maxiter': 500,\n 'disp': False\n }\n )\n\n best_h = result_slsqp.x\n best_c5 = result_slsqp.fun\n\n return (best_h, best_c5, n_points)\n```",
64 "env/all/time/policy": 493.082740210928,
65 "env/all/time/policy/min": 215.31174230575562,
66 "env/all/time/policy/max": 771.7152693271637,
67 "env/all/time/env_step": 978.2773189232685,
68 "env/all/time/env_step/min": 0.004937171936035156,
69 "env/all/time/env_step/max": 2562.9273443222046,
70 "env/all/time/reward_compute": 5.243346095085144e-07,
71 "env/all/time/reward_compute/min": 3.0919909477233887e-07,
72 "env/all/time/reward_compute/max": 1.0542571544647217e-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": 0.03739132359623909,
77 "advantage/min": -1.0,
78 "advantage/max": 9.992655754089355,
79 "time/assemble_training_data": 10.042886018753052,
80 "time/kl_vs_base": 132.2616832256317,
81 "kl_policy_base": 0.0008916512597352266,
82 "time/train": 989.4860379695892,
83 "time/save_checkpoint": 19.222825527191162,
84 "time/total": 4437.742107868195
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