Views
No views yet
erdos. Checkpoint saved
after training step 34 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 34,
3 "progress/batch": 34,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.7,
6 "puct/buffer_size": 551,
7 "puct/sampled_size": 8,
8 "puct/T": 17408,
9 "puct/scale_last": 0.1190591760561926,
10 "puct/buffer_value/mean": -0.3835226523578557,
11 "puct/buffer_value/std": 0.016229285002267963,
12 "puct/buffer_value/min": -0.5236859987110454,
13 "puct/buffer_value/max": -0.3809408239438133,
14 "puct/buffer_timestep/mean": 16.22141560798548,
15 "puct/buffer_timestep/std": 9.95315591642161,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 33.0,
18 "puct/buffer_construction_len/mean": 79.17785843920146,
19 "puct/buffer_construction_len/std": 9.452390000440758,
20 "puct/buffer_construction_len/min": 42.0,
21 "puct/buffer_construction_len/max": 143.0,
22 "puct/sampled_value/mean": -0.3809415052804609,
23 "puct/sampled_value/std": 4.787423147054087e-09,
24 "puct/sampled_value/min": -0.38094151114749253,
25 "puct/sampled_value/max": -0.380941498685385,
26 "puct/sampled_timestep/mean": 33.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 33.0,
29 "puct/sampled_timestep/max": 33.0,
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": 4654.226561784744,
35 "env/all/ac_tokens_per_turn": 9884.798828125,
36 "env/all/ob_tokens_per_turn": 1248.625,
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": 5061017,
41 "env/all/total_ob_tokens": 639296,
42 "env/all/time/sampling_mean": 317.90628521470353,
43 "env/all/time/sampling_max": 476.7936022281647,
44 "env/all/time/env_step_mean": 1976.9875669861212,
45 "env/all/time/env_step_max": 4225.109031677246,
46 "env/all/reward/mean": 0.5672901814111034,
47 "env/all/reward/max": 2.6250749180082527,
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": 0.5672901814111034,
53 "env/all/correctness": 0.216796875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.38242048490505803,
57 "env/all/raw_score/min": 0.3809414988079617,
58 "env/all/raw_score/max": 0.4631497956418942,
59 "env/all/initial_raw_score": -0.3809415052804609,
60 "env/all/initial_raw_score/min": -0.38094151114749253,
61 "env/all/initial_raw_score/max": -0.380941498685385,
62 "env/all/msg": "RuntimeError: Program execution failed: IndexError: arrays used as indices must be of integer (or boolean) type",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\nfrom math import inf\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n max_c5 = 0.0\n # Generate k values in the range [-2, 2]\n k_values = np.arange(-2.0, 2.0 + 1e-6, dx)\n for k in k_values:\n # Compute delta = k / dx\n delta = k / dx\n delta_rounded = np.round(delta)\n # Compute valid i range\n i_low = np.ceil((-k) / dx)\n i_high = np.floor((2 - k) / dx)\n i_low = max(0, i_low)\n i_high = min(n_points - 1, i_high)\n if i_low > i_high:\n continue\n # Get valid indices i in [i_low, i_high]\n i = np.arange(i_low, i_high + 1, dtype=int)\n j = i + delta_rounded\n j = np.clip(j, 0, n_points - 1)\n # Get valid indices for j\n valid_indices = (j >= 0) & (j < n_points)\n if np.any(valid_indices):\n integral = (h[i[valid_indices]] * (1.0 - h[j[valid_indices]])).sum() * dx\n max_c5 = max(max_c5, integral)\n return max_c5\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n import random\n random.seed(seed)\n\n if initial_h_values is not None:\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2\n initial_guess = initial_h_values.copy()\n else:\n n_points = 128 # Increased for better precision\n dx = 2.0 / n_points\n required_sum = n_points / 2\n # Binary initial guess\n indices = np.random.choice(n_points, size=int(required_sum), replace=False)\n initial_guess = np.zeros(n_points)\n initial_guess[indices] = 1.0\n\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 # Penalty for sum constraint\n penalty = 1e5 * (sum_h - required_sum) ** 2\n # Evaluate C5\n c5 = evaluate_c5(h, dx)\n return c5 + penalty\n\n # Adjusted parameters for better convergence\n result = differential_evolution(\n objective,\n bounds,\n strategy='rand1bin',\n popsize=100,\n maxiter=1000,\n tol=1e-9,\n mutation=(0.5, 1.0),\n recombination=0.9,\n seed=seed,\n x0=initial_guess\n )\n\n best_h = result.x\n best_c5 = evaluate_c5(best_h, dx)\n return best_h, best_c5, n_points\n```",
64 "env/all/time/policy": 317.90628521470353,
65 "env/all/time/policy/min": 147.13108801841736,
66 "env/all/time/policy/max": 476.7936022281647,
67 "env/all/time/env_step": 1976.9875669861212,
68 "env/all/time/env_step/min": 0.00699615478515625,
69 "env/all/time/env_step/max": 4225.109031677246,
70 "env/all/time/reward_compute": 3.7439167499542236e-07,
71 "env/all/time/reward_compute/min": 2.2724270820617676e-07,
72 "env/all/time/reward_compute/max": 9.201467037200928e-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": 0.04063553363084793,
77 "advantage/min": -1.0,
78 "advantage/max": 14.112588882446289,
79 "time/assemble_training_data": 8.332657098770142,
80 "time/kl_vs_base": 84.64232850074768,
81 "kl_policy_base": 0.0008370850118808448,
82 "time/train": 566.9060950279236,
83 "time/save_checkpoint": 17.930497646331787,
84 "time/total": 5333.299000740051
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