Views
No views yet
erdos. Checkpoint saved
after training step 6 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 6,
3 "progress/batch": 6,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.14,
6 "puct/buffer_size": 104,
7 "puct/sampled_size": 8,
8 "puct/T": 3072,
9 "puct/scale_last": 0.02418131231623838,
10 "puct/buffer_value/mean": -0.3923314668408866,
11 "puct/buffer_value/std": 0.03337415537289588,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38094965980730683,
14 "puct/buffer_timestep/mean": 2.230769230769231,
15 "puct/buffer_timestep/std": 1.8873606379054302,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 5.0,
18 "puct/buffer_construction_len/mean": 75.25961538461539,
19 "puct/buffer_construction_len/std": 21.13736901988609,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38097125138268584,
23 "puct/sampled_value/std": 1.8832419938802253e-05,
24 "puct/sampled_value/min": -0.3810008510962572,
25 "puct/sampled_value/max": -0.38094965980730683,
26 "puct/sampled_timestep/mean": 5.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 5.0,
29 "puct/sampled_timestep/max": 5.0,
30 "puct/sampled_construction_len/mean": 80.0,
31 "puct/sampled_construction_len/std": 0.0,
32 "puct/sampled_construction_len/min": 80.0,
33 "puct/sampled_construction_len/max": 80.0,
34 "time/sampling": 2703.8844022750854,
35 "env/all/ac_tokens_per_turn": 9009.8515625,
36 "env/all/ob_tokens_per_turn": 1324.75,
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": 4613044,
41 "env/all/total_ob_tokens": 678272,
42 "env/all/time/sampling_mean": 479.09659975534305,
43 "env/all/time/sampling_max": 705.3011329174042,
44 "env/all/time/env_step_mean": 644.4835897292942,
45 "env/all/time/env_step_max": 2023.1183524131775,
46 "env/all/reward/mean": 0.5979520341734815,
47 "env/all/reward/max": 2.6250433148950068,
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": 0.5979520341734815,
53 "env/all/correctness": 0.248046875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.42042669783320014,
57 "env/all/raw_score/min": 0.3809460849942446,
58 "env/all/raw_score/max": 0.5248879460400043,
59 "env/all/initial_raw_score": -0.3809712513826859,
60 "env/all/initial_raw_score/min": -0.3810008510962572,
61 "env/all/initial_raw_score/max": -0.38094965980730683,
62 "env/all/msg": "Computed C5 is not finite: nan",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import basinhopping\nfrom scipy.optimize import Bounds\n\ndef generate_perturbed_guesses(h, num_guesses=10, std=0.015, required_sum=None):\n \"\"\"Generate a list of perturbed guesses based on the given h.\"\"\"\n if required_sum is None:\n required_sum = h.sum()\n guesses = [h.copy()]\n for _ in range(num_guesses - 1):\n perturbed = h.copy() + np.random.normal(0, std, h.shape)\n perturbed = np.clip(perturbed, 0.0, 1.0)\n sum_perturbed = perturbed.sum()\n if sum_perturbed != 0:\n scale = required_sum / sum_perturbed\n perturbed = perturbed * scale\n guesses.append(perturbed)\n return guesses\n\ndef generate_random_initial(n_points, required_sum=None):\n \"\"\"Generate a random initial guess that satisfies the sum constraint.\"\"\"\n if required_sum is None:\n required_sum = n_points / 2.0\n h = np.random.rand(n_points)\n h = h / np.sum(h) * required_sum\n return h\n\ndef run(seed=42, budget_s=1000, **kwargs):\n initial_h = initial_h_values.copy()\n n_points = len(initial_h)\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n\n # Generate more diverse initial guesses\n initial_guesses = []\n\n # Perturbed guesses around initial_h_values\n for _ in range(10): # More perturbed guesses\n perturbed = initial_h.copy() + np.random.normal(0, 0.015, n_points)\n perturbed = np.clip(perturbed, 0.0, 1.0)\n sum_perturbed = perturbed.sum()\n if sum_perturbed != 0:\n scale = required_sum / sum_perturbed\n perturbed = perturbed * scale\n initial_guesses.append(perturbed)\n \n # Random guesses\n for _ in range(8): # More random guesses\n guess = generate_random_initial(n_points)\n initial_guesses.append(guess)\n\n best_h = None\n best_c5 = float('inf')\n\n # Objective function\n def objective(h):\n h = np.array(h)\n h1 = 1.0 - h\n corr = np.correlate(h, h1, mode='full')\n max_corr = np.max(corr)\n return max_corr * dx\n\n # Define bounds and constraints\n bounds = Bounds(0.0, 1.0)\n cons = {'type': 'eq', 'fun': lambda h: np.sum(h) - required_sum}\n\n # Run basinhopping with better parameters\n for guess in initial_guesses:\n result = basinhopping(\n objective,\n x0=guess,\n niter=500, # Increased iteration count\n T=0.05, # Slightly higher temperature\n stepsize=0.01, # Small step size for refinement\n minimizer_kwargs={\n 'method': 'L-BFGS-B', # More robust local optimizer\n 'bounds': bounds,\n 'constraints': cons,\n 'options': {'ftol': 1e-8, 'maxiter': 600}\n }\n )\n\n current_c5 = result.fun\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_h = result.x\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 479.09659975534305,
65 "env/all/time/policy/min": 204.47933197021484,
66 "env/all/time/policy/max": 705.3011329174042,
67 "env/all/time/env_step": 644.4835897292942,
68 "env/all/time/env_step/min": 0.006376981735229492,
69 "env/all/time/env_step/max": 2023.1183524131775,
70 "env/all/time/reward_compute": 6.07222318649292e-07,
71 "env/all/time/reward_compute/min": 2.086162567138672e-07,
72 "env/all/time/reward_compute/max": 2.4065375328063965e-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": 0.034374602138996124,
77 "advantage/min": -0.9043010473251343,
78 "advantage/max": 14.990449905395508,
79 "time/assemble_training_data": 5.667760372161865,
80 "time/kl_vs_base": 123.1171019077301,
81 "kl_policy_base": 0.0006966522778384387,
82 "time/train": 958.2604115009308,
83 "time/save_checkpoint": 13.517565250396729,
84 "time/total": 3805.9947233200073
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