Views
No views yet
erdos. Checkpoint saved
after training step 36 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 36,
3 "progress/batch": 36,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.74,
6 "puct/buffer_size": 584,
7 "puct/sampled_size": 8,
8 "puct/T": 18432,
9 "puct/scale_last": 0.2790759619298157,
10 "puct/buffer_value/mean": -0.3839368360408238,
11 "puct/buffer_value/std": 0.019326821317622936,
12 "puct/buffer_value/min": -0.66,
13 "puct/buffer_value/max": -0.3809240380701843,
14 "puct/buffer_timestep/mean": 17.246575342465754,
15 "puct/buffer_timestep/std": 10.538618897736114,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 35.0,
18 "puct/buffer_construction_len/mean": 94.27397260273973,
19 "puct/buffer_construction_len/std": 32.332713945286784,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 200.0,
22 "puct/sampled_value/mean": -0.38092501071000273,
23 "puct/sampled_value/std": 3.9247459429327167e-07,
24 "puct/sampled_value/min": -0.3809252561334897,
25 "puct/sampled_value/max": -0.3809240380701843,
26 "puct/sampled_timestep/mean": 35.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 35.0,
29 "puct/sampled_timestep/max": 35.0,
30 "puct/sampled_construction_len/mean": 100.0,
31 "puct/sampled_construction_len/std": 0.0,
32 "puct/sampled_construction_len/min": 100.0,
33 "puct/sampled_construction_len/max": 100.0,
34 "time/sampling": 1760.60369181633,
35 "env/all/ac_tokens_per_turn": 9753.689453125,
36 "env/all/ob_tokens_per_turn": 1642.125,
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": 4993889,
41 "env/all/total_ob_tokens": 840768,
42 "env/all/time/sampling_mean": 567.2069350276142,
43 "env/all/time/sampling_max": 829.6941080093384,
44 "env/all/time/env_step_mean": 18.95986255025491,
45 "env/all/time/env_step_max": 1100.1222214698792,
46 "env/all/reward/mean": 0.8776864650118265,
47 "env/all/reward/max": 2.6251952468299056,
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": 0.8776864650118265,
53 "env/all/correctness": 0.341796875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.3930200257967537,
57 "env/all/raw_score/min": 0.3809240379189368,
58 "env/all/raw_score/max": 1.0,
59 "env/all/initial_raw_score": -0.38092501071000273,
60 "env/all/initial_raw_score/min": -0.3809252561334897,
61 "env/all/initial_raw_score/max": -0.3809240380701843,
62 "env/all/msg": "RuntimeError: Program execution failed: ValueError: The number of bounds is not compatible with the length of `x0`.",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\nfrom scipy.signal import correlate\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 200\n dx = 2.0 / n_points\n desired_sum = n_points / 2\n\n # Use the provided initial sample or a more balanced one\n initial_h = np.array(initial_h_values, dtype=np.float64)\n initial_h = np.clip(initial_h, 0.0, 1.0)\n sum_h = np.sum(initial_h)\n if abs(desired_sum - sum_h) > 1e-6:\n initial_h[-1] += (desired_sum - sum_h)\n initial_h = np.clip(initial_h, 0.0, 1.0)\n\n # Set initial t\n initial_t = 0.38092\n x0 = np.concatenate((initial_h, [initial_t]))\n\n def objective(x):\n return x[-1]\n\n # Constraint: total integral of h(x) must be 1\n def constraint_sum(x):\n return np.sum(x[:-1]) - desired_sum\n\n # Compute overlaps for relevant k values only\n def compute_overlap(h, k):\n if k > 0:\n i_start = 0\n i_end = n_points - k - 1\n if i_start > i_end:\n return 0.0\n j = np.arange(i_start, i_end + 1) + k\n valid = (j >= 0) & (j < n_points)\n values = h[np.arange(i_start, i_end + 1)[valid]] * (1 - h[j[valid]])\n return np.sum(values) * dx\n elif k < 0:\n i_start = -k\n i_end = n_points - 1\n if i_start > i_end:\n return 0.0\n j = np.arange(i_start, i_end + 1) + k\n valid = (j >= 0) & (j < n_points)\n values = h[np.arange(i_start, i_end + 1)[valid]] * (1 - h[j[valid]])\n return np.sum(values) * dx\n else: # k == 0\n return np.sum(h * (1 - h)) * dx\n\n # Compute constraints for k = -2, -1, 0, 1, 2\n k_values = [-2, -1, 0, 1, 2]\n\n def constraints_func(x):\n h = x[:-1]\n t = x[-1]\n overlaps = [compute_overlap(h, k) for k in k_values]\n return [t - overlap for overlap in overlaps]\n\n cons = [\n {'type': 'eq', 'fun': constraint_sum},\n {'type': 'ineq', 'fun': constraints_func}\n ]\n\n bounds = [(0.0, 1.0) for _ in range(n_points)] + [(0.0, np.inf)]\n\n result = minimize(\n fun=objective,\n x0=x0,\n bounds=bounds,\n constraints=cons,\n method='SLSQP',\n options={\n 'ftol': 1e-9,\n 'maxiter': 5000,\n 'disp': False\n }\n )\n\n if result.success:\n best_h = result.x[:-1]\n best_t = result.x[-1]\n else:\n best_h = initial_h.copy()\n best_t = initial_t\n\n return best_h, best_t, n_points\n```",
64 "env/all/time/policy": 567.2069350276142,
65 "env/all/time/policy/min": 256.6535475254059,
66 "env/all/time/policy/max": 829.6941080093384,
67 "env/all/time/env_step": 18.95986255025491,
68 "env/all/time/env_step/min": 0.005200862884521484,
69 "env/all/time/env_step/max": 1100.1222214698792,
70 "env/all/time/reward_compute": 5.634501576423645e-07,
71 "env/all/time/reward_compute/min": 2.4959444999694824e-07,
72 "env/all/time/reward_compute/max": 2.2761523723602295e-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": 0.02587207406759262,
77 "advantage/min": -1.0,
78 "advantage/max": 6.506897449493408,
79 "time/assemble_training_data": 11.028441429138184,
80 "time/kl_vs_base": 146.56563568115234,
81 "kl_policy_base": 0.0008501674165017903,
82 "time/train": 1078.959444284439,
83 "time/save_checkpoint": 18.40455985069275,
84 "time/total": 3017.4046082496643
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