Views
No views yet
erdos. Checkpoint saved
after training step 1 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 1,
3 "progress/batch": 1,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.04,
6 "puct/buffer_size": 24,
7 "puct/sampled_size": 8,
8 "puct/T": 512,
9 "puct/scale_last": 0.022369287552525763,
10 "puct/buffer_value/mean": -0.4281021665929304,
11 "puct/buffer_value/std": 0.056234886766976216,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38276168457101944,
14 "puct/buffer_timestep/mean": -0.3333333333333333,
15 "puct/buffer_timestep/std": 0.4714045207910317,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 0.0,
18 "puct/buffer_construction_len/mean": 73.875,
19 "puct/buffer_construction_len/std": 23.0458103567655,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38345961695175057,
23 "puct/sampled_value/std": 0.0004664039038100035,
24 "puct/sampled_value/min": -0.3840456494138304,
25 "puct/sampled_value/max": -0.38276168457101944,
26 "puct/sampled_timestep/mean": 0.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 0.0,
29 "puct/sampled_timestep/max": 0.0,
30 "puct/sampled_construction_len/mean": 80.125,
31 "puct/sampled_construction_len/std": 29.07506448832057,
32 "puct/sampled_construction_len/min": 42.0,
33 "puct/sampled_construction_len/max": 143.0,
34 "time/sampling": 1822.9598400592804,
35 "env/all/ac_tokens_per_turn": 9603.720703125,
36 "env/all/ob_tokens_per_turn": 1166.125,
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": 4917105,
41 "env/all/total_ob_tokens": 597056,
42 "env/all/time/sampling_mean": 528.5299022924155,
43 "env/all/time/sampling_max": 753.7376616001129,
44 "env/all/time/env_step_mean": 20.07542957831174,
45 "env/all/time/env_step_max": 1100.1098220348358,
46 "env/all/reward/mean": 0.8788887084099563,
47 "env/all/reward/max": 2.6178186126275564,
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": 0.8788887084099563,
53 "env/all/correctness": 0.361328125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.4179968519034591,
57 "env/all/raw_score/min": 0.3819974267881357,
58 "env/all/raw_score/max": 0.993006993006993,
59 "env/all/initial_raw_score": -0.3834596169517505,
60 "env/all/initial_raw_score/min": -0.3840456494138304,
61 "env/all/initial_raw_score/max": -0.38276168457101944,
62 "env/all/msg": "RuntimeError: Program execution failed: TypeError: differential_evolution() got an unexpected keyword argument 'fun'",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\n\ndef run(seed=42, budget_s=1000, **kwargs):\n # Choose number of points (42 was shown to perform well)\n n_points = 42\n dx = 2.0 / n_points\n\n # Initial guess is from the provided function; normalize if needed\n h = np.array(initial_h_values, dtype=np.float64)\n if np.sum(h) != n_points / 2:\n h = h * (n_points / 2) / np.sum(h)\n # We\u2019ll use this as a guide, but not directly as initial population\n\n # Bounds for the optimization variables\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Objective function: maximize overlap; we minimize max_corr\n def objective(y):\n # Normalize and ensure the sum constraint is satisfied\n sum_y = np.sum(y)\n if sum_y == 0:\n return np.inf # Avoid division by zero\n h = (y / sum_y) * (n_points / 2)\n h = np.clip(h, 0, 1) # Ensure values stay in [0, 1]\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n max_corr = np.max(corr) * dx\n # Penalties for constraint violations\n penalty = 1e4 * np.abs(np.sum(h) - n_points / 2)\n return max_corr + penalty # Total objective to minimize\n\n # Run Differential Evolution\n result = differential_evolution(\n fun=objective,\n bounds=bounds,\n strategy='best1bin',\n maxiter=200,\n popsize=30,\n tol=1e-6,\n mutation=(0.5, 1),\n recombination=0.8\n )\n\n best_h = result.x / np.sum(result.x) * (n_points / 2)\n best_h = np.clip(best_h, 0, 1)\n # Re-calculate the actual max_corr to exclude any small penalty\n best_c5 = objective(result.x) - 1e4 * np.abs(np.sum(best_h) - n_points / 2)\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 528.5299022924155,
65 "env/all/time/policy/min": 268.77983951568604,
66 "env/all/time/policy/max": 753.7376616001129,
67 "env/all/time/env_step": 20.07542957831174,
68 "env/all/time/env_step/min": 0.005690336227416992,
69 "env/all/time/env_step/max": 1100.1098220348358,
70 "env/all/time/reward_compute": 6.142072379589081e-07,
71 "env/all/time/reward_compute/min": 1.9371509552001953e-07,
72 "env/all/time/reward_compute/max": 1.866370439529419e-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": 0.024153785780072212,
77 "advantage/min": -0.9935250282287598,
78 "advantage/max": 2.9061083793640137,
79 "time/assemble_training_data": 9.52767539024353,
80 "time/kl_vs_base": 133.5515534877777,
81 "kl_policy_base": 0.000545418937690556,
82 "time/train": 1004.8986876010895,
83 "time/save_checkpoint": 8.101783275604248,
84 "time/total": 2982.3145105838776
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