Views
No views yet
ac2. Checkpoint saved
after training step 29 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 29,
3 "progress/batch": 29,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.6,
6 "puct/buffer_size": 470,
7 "puct/sampled_size": 8,
8 "puct/T": 14848,
9 "puct/scale_last": 0.4399073664015871,
10 "puct/buffer_value/mean": 0.930250447205922,
11 "puct/buffer_value/std": 0.04475720452576275,
12 "puct/buffer_value/min": 0.5041471954736918,
13 "puct/buffer_value/max": 0.9440545618752789,
14 "puct/buffer_timestep/mean": 13.725531914893617,
15 "puct/buffer_timestep/std": 8.530245764464992,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 28.0,
18 "puct/buffer_construction_len/mean": 2730.531914893617,
19 "puct/buffer_construction_len/std": 2534.9247196096026,
20 "puct/buffer_construction_len/min": 1024.0,
21 "puct/buffer_construction_len/max": 32768.0,
22 "puct/sampled_value/mean": 0.9438997024614514,
23 "puct/sampled_value/std": 6.487951436738853e-05,
24 "puct/sampled_value/min": 0.9438578317072384,
25 "puct/sampled_value/max": 0.9440545618752789,
26 "puct/sampled_timestep/mean": 28.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 28.0,
29 "puct/sampled_timestep/max": 28.0,
30 "puct/sampled_construction_len/mean": 4096.0,
31 "puct/sampled_construction_len/std": 0.0,
32 "puct/sampled_construction_len/min": 4096.0,
33 "puct/sampled_construction_len/max": 4096.0,
34 "time/sampling": 4174.056032657623,
35 "env/all/ac_tokens_per_turn": 9762.16796875,
36 "env/all/ob_tokens_per_turn": 4037.25,
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": 4998230,
41 "env/all/total_ob_tokens": 2067072,
42 "env/all/time/sampling_mean": 401.1490901764482,
43 "env/all/time/sampling_max": 523.8003692626953,
44 "env/all/time/env_step_mean": 1650.9465017118491,
45 "env/all/time/env_step_max": 3650.1325421333313,
46 "env/all/reward/mean": 0.6823510093438542,
47 "env/all/reward/max": 0.94410944427093,
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": 0.6823510093438542,
53 "env/all/correctness": 0.74609375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.9145647036231764,
57 "env/all/raw_score/min": 0.026960664752654028,
58 "env/all/raw_score/max": 0.94410944427093,
59 "env/all/initial_raw_score": 0.9438997024614514,
60 "env/all/initial_raw_score/min": 0.9438578317072384,
61 "env/all/initial_raw_score/max": 0.9440545618752789,
62 "env/all/msg": "RuntimeError: Program execution failed: ValueError: Sum of sequence is too close to zero.",
63 "env/all/parsed_code": "```python\nimport numpy as np\nimport time\nfrom typing import Tuple\nimport random\nfrom scipy.optimize import minimize\n\ndef _simpson_l2sq(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n \"\"\"Compute ||f*f||_2^2 via Simpson's rule with endpoint zeros and gradient.\"\"\"\n m = conv.size\n if m == 0:\n return 0.0, np.zeros_like(conv)\n dx = 1.0 / (m + 1)\n y = np.zeros(m + 2, dtype=conv.dtype)\n y[0] = 0.0\n y[1:-1] = conv\n y[-1] = 0.0\n lhs = y[:-1]\n rhs = y[1:]\n l2_sq = (dx / 3.0) * np.sum(lhs * lhs + lhs * rhs + rhs * rhs)\n grad_y = (dx / 3.0) * (4.0 * y + np.roll(y, 1) + np.roll(y, -1))\n grad_conv = grad_y[1:-1]\n return float(l2_sq), grad_conv\n\ndef _l1(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n \"\"\"Compute ||f*f||_1 and its gradient.\"\"\"\n m = conv.size\n dx = 1.0 / (m + 1) if m > 0 else 1.0\n val = dx * float(np.sum(conv)) if m > 0 else 0.0\n grad = np.full_like(conv, dx)\n return val, grad\n\ndef _linf(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n \"\"\"Compute ||f*f||_inf and its subgradient.\"\"\"\n if conv.size == 0:\n return 0.0, np.zeros_like(conv)\n m = float(np.max(conv))\n mask = conv == m\n count = int(mask.sum())\n if count == 0 or m <= 0.0:\n return m, np.zeros_like(conv)\n grad = mask.astype(conv.dtype)\n return m, grad\n\ndef _objective_and_grad_conv(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n \"\"\"Compute C = l2_sq / (l1 * linf) and its gradient.\"\"\"\n l2_sq, g_l2 = _simpson_l2sq(conv)\n l1, g_l1 = _l1(conv)\n linf, g_linf = _linf(conv)\n if l1 <= 0.0 or linf <= 0.0:\n return 0.0, np.zeros_like(conv)\n denom = l1 * linf\n c_value = l2_sq / denom\n num_grad = g_l2 * denom - l2_sq * (g_l1 * linf + l1 * g_linf)\n g_conv = num_grad / (denom * denom)\n return float(c_value), g_conv\n\ndef _grad_h_from_conv_grad(h: np.ndarray, g_conv: np.ndarray) -> np.ndarray:\n \"\"\"Compute gradient of C w.r.t h from gradient of C w.r.t conv.\"\"\"\n h_rev = h[::-1]\n g_h = np.convolve(g_conv, h_rev, mode=\"full\")\n N = len(h)\n gradient = g_h[N-1 : N-1 + N]\n return gradient\n\ndef construct_function():\n \"\"\"Optimize step function sequence using a combination of structured initializations and guided optimization.\"\"\"\n print(\"Starting optimized sequence construction.\")\n max_seconds = 1000\n n_start = 128\n \n # Initialize from previous best if available\n initial_sequences = []\n if 'height_sequence_1' in globals():\n initial_sequences.append(np.array(height_sequence_1, dtype=np.float32))\n else:\n pass # Proceed with generated sequences\n\n # Generate diverse structured initial sequences with guaranteed sum of 0.01\n for _ in range(32):\n # Flat sequence with sum 0.01\n seq = np.full(n_start, (0.01) / n_start, dtype=np.float32)\n initial_sequences.append(seq)\n\n # Single peak with sum 0.01\n seq = np.zeros(n_start, dtype=np.float32)\n seq[n_start // 2] = 0.01\n initial_sequences.append(seq)\n\n # Sinusoidal pattern with sum 0.01\n seq = np.sin(np.linspace(0, 2 * np.pi, n_start))\n seq = np.clip(seq, 0.0, 1.0)\n sum_val = np.sum(seq)\n if sum_val > 0:\n seq = seq * (0.01) / sum_val\n else:\n seq = np.full(n_start, (0.01) / n_start)\n initial_sequences.append(seq)\n\n # Gaussian-like pattern with sum 0.01\n seq = np.zeros(n_start, dtype=np.float32)\n center = n_start // 2\n std = n_start // 8\n for i in range(n_start):\n seq[i] = np.exp(-((i - center) ** 2) / (2 * std ** 2))\n sum_val = np.sum(seq)\n if sum_val > 0:\n seq = seq * (0.01) / sum_val\n else:\n seq = np.full(n_start, (0.01) / n_start)\n initial_sequences.append(seq)\n\n # Random sparse peaks\n seq = np.zeros(n_start, dtype=np.float32)\n num_sparse_peaks = 5\n indices = np.random.choice(n_start, num_sparse_peaks, replace=False)\n seq[indices] = 0.01 / num_sparse_peaks\n initial_sequences.append(seq)\n\n # Evaluate initial sequences to find the best one\n best_seq = None\n best_value = -float('inf')\n for seq in initial_sequences:\n try:\n current_value = evaluate_sequence(seq.tolist())\n except Exception as e:\n print(f\"Error evaluating initial sequence: {e}\")\n continue\n if current_value > best_value:\n best_value = current_value\n best_seq = seq.copy()\n\n if best_seq is None:\n # Fallback to random initialization with sum 0.01\n best_seq = np.random.rand(n_start) * 1000.0\n best_seq = np.clip(best_seq, 0.0, 1000.0)\n best_seq = best_seq * (0.01 / np.sum(best_seq))\n best_value = evaluate_sequence(best_seq.tolist())\n\n # Try different lengths for sequences\n target_lengths = [256, 512, 1024, 4096]\n best_result = best_value\n best_seq_resized = best_seq.copy()\n\n for length in target_lengths:\n if len(best_seq) < length:\n upsample_factor = (length + len(best_seq) - 1) // len(best_seq)\n upsampled_seq = np.repeat(best_seq, upsample_factor, axis=0)\n upsampled_seq = upsampled_seq[:length]\n else:\n upsampled_seq = best_seq[:length]\n \n # Ensure sum is exactly 0.01\n sum_val = np.sum(upsampled_seq)\n if sum_val > 0:\n upsampled_seq = upsampled_seq * (0.01 / sum_val)\n else:\n upsampled_seq = np.full(length, (0.01) / length)\n \n # Evaluate\n current_value = evaluate_sequence(upsampled_seq.tolist())\n if current_value > best_value:\n best_value = current_value\n best_seq_resized = upsampled_seq.copy()\n \n # Initial candidate\n best_seq = best_seq_resized.copy()\n\n start_time = time.time()\n\n # Define the objective function for optimization\n def objective(x):\n seq = x.copy()\n sum_val = np.sum(seq)\n if sum_val > 0:\n seq = seq * (0.01 / sum_val)\n else:\n seq = np.full(len(seq), (0.01) / len(seq))\n return -evaluate_sequence(seq.tolist()) # Minimize negative to maximize\n\n # Define bounds and constraints\n bounds = [(0.0, 1000.0) for _ in range(len(best_seq))]\n constraints = [{'type': 'ineq', 'fun': lambda x: np.sum(x) - 0.01}]\n \n # Use L-BFGS-B optimizer for local refinement\n result = minimize(objective, best_seq, method='L-BFGS-B', bounds=bounds, constraints=constraints, jac=None, tol=1e-6)\n \n final_seq = result.x\n final_value = -objective(final_seq)\n print(f\"Final C2 lower bound: {final_value:.6f}\")\n return final_seq.tolist()\n```",
64 "env/all/time/policy": 401.1490901764482,
65 "env/all/time/policy/min": 186.3613317012787,
66 "env/all/time/policy/max": 523.8003692626953,
67 "env/all/time/env_step": 1650.9465017118491,
68 "env/all/time/env_step/min": 0.006981611251831055,
69 "env/all/time/env_step/max": 3650.1325421333313,
70 "env/all/time/reward_compute": 2.3469328880310059e-07,
71 "env/all/time/reward_compute/min": 1.8998980522155762e-07,
72 "env/all/time/reward_compute/max": 3.46451997756958e-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": 0.017478201538324356,
77 "advantage/min": -1.0,
78 "advantage/max": 1.7320470809936523,
79 "time/assemble_training_data": 11.085632085800171,
80 "time/kl_vs_base": 101.33315086364746,
81 "kl_policy_base": 0.0006670355214737356,
82 "time/train": 681.8683898448944,
83 "time/save_checkpoint": 13.931088924407959,
84 "time/total": 4986.549501657486
85}[2026-07-09T06:38:34+00:00] job=1812632 node=node-30 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T06:45:52+00:00] job=1812704 node=node-30 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:00:42+00:00] job=1812735 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:26:33+00:00] job=1812827 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:21:09+00:00] job=1813131 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T14:53:48+00:00] job=1813132 node=node-2 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-10T03:31:51+00:00] job=1816627 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-10T03:54:03+00:00] job=1816628 node=node-29 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-10T07:56:44+00:00] job=1817463 node=node-27 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-10T09:10:26+00:00] job=1817464 node=node-7 ngpu=6 ntrain=2 replicas=4 flash_attn=yes