Views
No views yet
erdos. Checkpoint saved
after training step 8 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 8,
3 "progress/batch": 8,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.18,
6 "puct/buffer_size": 136,
7 "puct/sampled_size": 8,
8 "puct/T": 4096,
9 "puct/scale_last": 0.030362440548684477,
10 "puct/buffer_value/mean": -0.38970295629089563,
11 "puct/buffer_value/std": 0.02799683379898324,
12 "puct/buffer_value/min": -0.5130522804051018,
13 "puct/buffer_value/max": -0.38101721320875104,
14 "puct/buffer_timestep/mean": 3.235294117647059,
15 "puct/buffer_timestep/std": 2.462170533700747,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 7.0,
18 "puct/buffer_construction_len/mean": 61.779411764705884,
19 "puct/buffer_construction_len/std": 21.16972421132101,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 128.0,
22 "puct/sampled_value/mean": -0.38110812451250076,
23 "puct/sampled_value/std": 6.678137296835784e-05,
24 "puct/sampled_value/min": -0.38117816952344685,
25 "puct/sampled_value/max": -0.38101721320875104,
26 "puct/sampled_timestep/mean": 7.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 7.0,
29 "puct/sampled_timestep/max": 7.0,
30 "puct/sampled_construction_len/mean": 86.5,
31 "puct/sampled_construction_len/std": 18.103866990231673,
32 "puct/sampled_construction_len/min": 64.0,
33 "puct/sampled_construction_len/max": 128.0,
34 "time/sampling": 1610.1571815013885,
35 "env/all/ac_tokens_per_turn": 8835.486328125,
36 "env/all/ob_tokens_per_turn": 1430.5,
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": 4523769,
41 "env/all/total_ob_tokens": 732416,
42 "env/all/time/sampling_mean": 475.1154082301073,
43 "env/all/time/sampling_max": 678.5025298595428,
44 "env/all/time/env_step_mean": 81.56391004938632,
45 "env/all/time/env_step_max": 1100.127678155899,
46 "env/all/reward/mean": 0.5763518380640373,
47 "env/all/reward/max": 2.6246662407154275,
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": 0.5763518380640373,
53 "env/all/correctness": 0.2265625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.39611926010237386,
57 "env/all/raw_score/min": 0.3810008138332892,
58 "env/all/raw_score/max": 0.5183051091918132,
59 "env/all/initial_raw_score": -0.38110812451250087,
60 "env/all/initial_raw_score/min": -0.38117816952344685,
61 "env/all/initial_raw_score/max": -0.38101721320875104,
62 "env/all/msg": "Evaluation timed out after 1100 seconds.",
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 = 64 # Use a moderate number of points for a balance between accuracy and runtime\n dx = 2.0 / n_points\n\n # Start with initial_h_values and generate additional guesses\n initial_guess = initial_h_values.copy()\n\n # Generate multiple initial guesses by perturbing the initial_h_values\n # and ensuring the constraint is enforced\n def generate_perturbation(guess):\n perturbation = np.random.uniform(-0.05, 0.05, n_points)\n perturbed = guess + perturbation\n sum_perturbed = np.sum(perturbed)\n if sum_perturbed != 0:\n perturbed = perturbed * (n_points / 2) / sum_perturbed\n return np.clip(perturbed, 0.0, 1.0)\n\n # Create a list of initial guesses\n initial_guesses = [initial_guess.copy()]\n for _ in range(15): # Add 15 perturbed variants to the initial guesses\n guess = generate_perturbation(initial_guess)\n initial_guesses.append(guess)\n\n best_c5 = float('inf')\n best_h = np.zeros(n_points)\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Local optimizer used for refinement\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(h):\n return np.sum(h) - n_points / 2\n\n for guess in initial_guesses:\n # Basin-hopping with more aggressive search parameters\n result_bh = basinhopping(\n objective,\n guess,\n niter=200, # More global search iterations\n T=0.1, # Lower temperature to reduce noise\n stepsize=0.1,\n minimizer_kwargs={\n 'method': 'SLSQP',\n 'bounds': bounds,\n 'constraints': [{'type': 'eq', 'fun': constraint}],\n 'options': {'ftol': 1e-12, 'maxiter': 500} # Tighter tolerance\n },\n seed=seed\n )\n\n # Final local refinement\n result_slsqp = minimize(\n fun=objective,\n x0=result_bh.x,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint}],\n options={'ftol': 1e-12, 'maxiter': 500}\n )\n\n # Update if a better solution is found\n if result_slsqp.fun < best_c5:\n best_c5 = result_slsqp.fun\n best_h = result_slsqp.x\n\n # Perturb and refine further for better convergence\n for _ in range(50): # More aggressive fine-tuning\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 sum_perturbed = np.sum(h_perturbed)\n h_perturbed = h_perturbed * (n_points / 2) / sum_perturbed\n result_refine = minimize(\n fun=objective,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint}],\n options={'ftol': 1e-12, 'maxiter': 200}\n )\n if result_refine.fun < best_c5:\n best_c5 = result_refine.fun\n best_h = result_refine.x\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 475.1154082301073,
65 "env/all/time/policy/min": 163.80647921562195,
66 "env/all/time/policy/max": 678.5025298595428,
67 "env/all/time/env_step": 81.56391004938632,
68 "env/all/time/env_step/min": 0.005884885787963867,
69 "env/all/time/env_step/max": 1100.127678155899,
70 "env/all/time/reward_compute": 3.348104655742645e-07,
71 "env/all/time/reward_compute/min": 2.7194619178771973e-07,
72 "env/all/time/reward_compute/max": 5.066394805908203e-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": 0.03573092073202133,
77 "advantage/min": -0.9026234745979309,
78 "advantage/max": 14.149524688720703,
79 "time/assemble_training_data": 5.427765369415283,
80 "time/kl_vs_base": 125.65817952156067,
81 "kl_policy_base": 0.0006905627087689936,
82 "time/train": 942.0744025707245,
83 "time/save_checkpoint": 11.00989580154419,
84 "time/total": 2695.398871421814
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