Views
No views yet
erdos. Checkpoint saved
after training step 27 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 27,
3 "progress/batch": 27,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.56,
6 "puct/buffer_size": 440,
7 "puct/sampled_size": 8,
8 "puct/T": 13824,
9 "puct/scale_last": 0.11905859715763034,
10 "puct/buffer_value/mean": -0.3841676667498393,
11 "puct/buffer_value/std": 0.018103980335595696,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.38094140284237554,
14 "puct/buffer_timestep/mean": 12.745454545454546,
15 "puct/buffer_timestep/std": 7.941188786248755,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 26.0,
18 "puct/buffer_construction_len/mean": 78.97045454545454,
19 "puct/buffer_construction_len/std": 10.567597979962887,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.38094152133552106,
23 "puct/sampled_value/std": 4.4816007722526484e-08,
24 "puct/sampled_value/min": -0.38094154019042364,
25 "puct/sampled_value/max": -0.38094140284237554,
26 "puct/sampled_timestep/mean": 26.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 26.0,
29 "puct/sampled_timestep/max": 26.0,
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": 4618.32190656662,
35 "env/all/ac_tokens_per_turn": 9350.205078125,
36 "env/all/ob_tokens_per_turn": 1434.375,
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": 4787305,
41 "env/all/total_ob_tokens": 734400,
42 "env/all/time/sampling_mean": 291.7911006179638,
43 "env/all/time/sampling_max": 412.7567162513733,
44 "env/all/time/env_step_mean": 1876.898315534927,
45 "env/all/time/env_step_max": 4219.493016004562,
46 "env/all/reward/mean": 0.2613218075509415,
47 "env/all/reward/max": 2.625076476821646,
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": 0.2613218075509415,
53 "env/all/correctness": 0.099609375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.38117846861823174,
57 "env/all/raw_score/min": 0.3809412725986564,
58 "env/all/raw_score/max": 0.38860765551528154,
59 "env/all/initial_raw_score": -0.38094152133552117,
60 "env/all/initial_raw_score/min": -0.38094154019042364,
61 "env/all/initial_raw_score/max": -0.38094140284237554,
62 "env/all/msg": "Evaluation timed out after 1100 seconds.",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\nimport random\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n # Valid shifts correspond to [n_points - 1 : 2 * n_points - 1]\n corr_valid = corr[n_points - 1 : 2 * n_points - 1]\n max_c5 = np.max(corr_valid) * dx\n return max_c5\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n random.seed(seed)\n \n # Use initial_h_values if available\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2\n \n # Prepare initial guess\n if initial_h_values is not None:\n initial_guess = initial_h_values.copy()\n else:\n # Randomly generate a valid initial guess\n initial_guess = np.random.rand(n_points)\n initial_guess = initial_guess / np.sum(initial_guess) * required_sum\n\n # Define bounds [0, 1]\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(x):\n h = np.array(x)\n sum_h = np.sum(h)\n # Reduced penalty coefficient\n penalty = 1e2 * (sum_h - required_sum) ** 2\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Optimization settings\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin', # Strategy to focus on exploration\n popsize=200, # Increase population for diversity\n maxiter=500, # More iterations\n tol=1e-6,\n mutation=(0.5, 1.0), # Wider mutation range for exploration\n recombination=0.9, # Higher recombination rate for diversity\n seed=seed,\n x0=initial_guess\n )\n\n best_h = result.x\n best_c5 = evaluate_c5(best_h, dx)\n\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 291.7911006179638,
65 "env/all/time/policy/min": 140.35489177703857,
66 "env/all/time/policy/max": 412.7567162513733,
67 "env/all/time/env_step": 1876.898315534927,
68 "env/all/time/env_step/min": 0.00533294677734375,
69 "env/all/time/env_step/max": 4219.493016004562,
70 "env/all/time/reward_compute": 3.157183527946472e-07,
71 "env/all/time/reward_compute/min": 2.3096799850463867e-07,
72 "env/all/time/reward_compute/max": 5.587935447692871e-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": 0.05499803274869919,
77 "advantage/min": -0.731345534324646,
78 "advantage/max": 14.112621307373047,
79 "time/assemble_training_data": 8.572037935256958,
80 "time/kl_vs_base": 78.79961895942688,
81 "kl_policy_base": 0.000876589328981936,
82 "time/train": 546.0702295303345,
83 "time/save_checkpoint": 17.02359366416931,
84 "time/total": 5270.377825021744
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
[2026-07-10T08:20:19+00:00] job=1813610 node=node-29 ngpu=6 ntrain=2 replicas=4 flash_attn=yes