Views
No views yet
ac2. Checkpoint saved
after training step 45 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 45,
3 "progress/batch": 45,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.92,
6 "puct/buffer_size": 726,
7 "puct/sampled_size": 8,
8 "puct/T": 23040,
9 "puct/scale_last": 0.4406305761214052,
10 "puct/buffer_value/mean": 0.9352590442992436,
11 "puct/buffer_value/std": 0.036645940047358126,
12 "puct/buffer_value/min": 0.5041471954736918,
13 "puct/buffer_value/max": 0.944777771595097,
14 "puct/buffer_timestep/mean": 21.75619834710744,
15 "puct/buffer_timestep/std": 13.153043154537325,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 44.0,
18 "puct/buffer_construction_len/mean": 3212.0192837465565,
19 "puct/buffer_construction_len/std": 2141.402991260169,
20 "puct/buffer_construction_len/min": 1024.0,
21 "puct/buffer_construction_len/max": 32768.0,
22 "puct/sampled_value/mean": 0.9447735397491113,
23 "puct/sampled_value/std": 3.2202330926389905e-06,
24 "puct/sampled_value/min": 0.9447679880761147,
25 "puct/sampled_value/max": 0.944777771595097,
26 "puct/sampled_timestep/mean": 44.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 44.0,
29 "puct/sampled_timestep/max": 44.0,
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": 5389.306136608124,
35 "env/all/ac_tokens_per_turn": 9857.853515625,
36 "env/all/ob_tokens_per_turn": 4460.875,
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": 5047221,
41 "env/all/total_ob_tokens": 2283968,
42 "env/all/time/sampling_mean": 725.0600522314198,
43 "env/all/time/sampling_max": 1030.7621109485626,
44 "env/all/time/env_step_mean": 2303.706905388739,
45 "env/all/time/env_step_max": 4388.473441362381,
46 "env/all/reward/mean": 0.8454962035141662,
47 "env/all/reward/max": 0.9447862533369672,
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": 0.8454962035141662,
53 "env/all/correctness": 0.90234375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.9370001216434051,
57 "env/all/raw_score/min": 0.44466145833333326,
58 "env/all/raw_score/max": 0.9447862533369672,
59 "env/all/initial_raw_score": 0.9447735397491113,
60 "env/all/initial_raw_score/min": 0.9447679880761147,
61 "env/all/initial_raw_score/max": 0.944777771595097,
62 "env/all/msg": "Success; raw_score=0.9447668922558193",
63 "env/all/parsed_code": "```python\nimport numpy as np\nimport time\nfrom typing import Tuple\nimport random\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 \"\"\"\n Construct a sequence of non-negative heights for a step function to maximize the evaluation function.\n Uses a combination of diverse initialization, adaptive upsampling, and optimized gradient ascent\n with constraint maintenance and improved hyperparameters.\n \"\"\"\n # Parameters for the algorithm\n n_start = 128\n target_length = 4096\n max_seconds = 1000\n num_initializations = 16\n\n # Initialize from available height_sequence_1 if it exists\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 initial sequences\n for _ in range(num_initializations):\n # Gaussian-like sequence\n seq = np.zeros(n_start, dtype=np.float32)\n center = n_start // 2\n std = n_start // 6\n for i in range(n_start):\n seq[i] = np.exp(-((i - center) ** 2) / (2 * std**2))\n # First scale to max constraint\n max_initial = np.max(seq)\n if max_initial > 1000.0:\n scale_factor = 1000.0 / max_initial\n seq = seq * scale_factor\n # Now scale to sum 0.01\n sum_initial = np.sum(seq)\n if sum_initial == 0:\n sum_initial = 1.0 # Avoid division by zero\n scale_factor = 0.01 / sum_initial\n seq = seq * scale_factor\n initial_sequences.append(seq)\n\n # Random sequence\n seq_rand = np.random.rand(n_start) * 0.1\n # First scale to max constraint\n max_rand = np.max(seq_rand)\n if max_rand > 1000.0:\n scale_factor = 1000.0 / max_rand\n seq_rand = seq_rand * scale_factor\n # Now scale to sum 0.01\n sum_rand = np.sum(seq_rand)\n if sum_rand == 0:\n sum_rand = 1.0\n scale_factor = 0.01 / sum_rand\n seq_rand = seq_rand * scale_factor\n initial_sequences.append(seq_rand)\n\n # Uniform sequence\n seq_uniform = np.ones(n_start, dtype=np.float32) * (0.01 / n_start)\n # First scale to max constraint\n max_uniform = np.max(seq_uniform)\n if max_uniform > 1000.0:\n scale_factor = 1000.0 / max_uniform\n seq_uniform = seq_uniform * scale_factor\n # Now scale to sum 0.01\n sum_uniform = np.sum(seq_uniform)\n if sum_uniform == 0:\n sum_uniform = 1.0\n scale_factor = 0.01 / sum_uniform\n seq_uniform = seq_uniform * scale_factor\n initial_sequences.append(seq_uniform)\n\n # Multi-peak sequence\n seq_multi = np.zeros(n_start, dtype=np.float32)\n peak_positions = [n_start // 4, n_start // 2, n_start * 3 // 4]\n for pos in peak_positions:\n seq_multi[pos] += 0.01 / len(peak_positions)\n # First scale to max constraint\n max_multi = np.max(seq_multi)\n if max_multi > 1000.0:\n scale_factor = 1000.0 / max_multi\n seq_multi = seq_multi * scale_factor\n # Now scale to sum 0.01\n sum_multi = np.sum(seq_multi)\n if sum_multi == 0:\n sum_multi = 1.0\n scale_factor = 0.01 / sum_multi\n seq_multi = seq_multi * scale_factor\n initial_sequences.append(seq_multi)\n\n # Single-peak sequence\n seq_single = np.zeros(n_start, dtype=np.float32)\n seq_single[n_start // 2] = 0.01\n # First scale to max constraint\n max_single = np.max(seq_single)\n if max_single > 1000.0:\n scale_factor = 1000.0 / max_single\n seq_single = seq_single * scale_factor\n # Now scale to sum 0.01\n sum_single = np.sum(seq_single)\n if sum_single == 0:\n sum_single = 1.0\n scale_factor = 0.01 / sum_single\n seq_single = seq_single * scale_factor\n initial_sequences.append(seq_single)\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 a Gaussian-like sequence\n seq = np.zeros(n_start, dtype=np.float32)\n center = n_start // 2\n std = n_start // 6\n for i in range(n_start):\n seq[i] = np.exp(-((i - center) ** 2) / (2 * std**2))\n # First scale to max constraint\n max_initial = np.max(seq)\n if max_initial > 1000.0:\n scale_factor = 1000.0 / max_initial\n seq = seq * scale_factor\n # Now scale to sum 0.01\n sum_initial = np.sum(seq)\n if sum_initial == 0:\n sum_initial = 1.0 # Avoid division by zero\n scale_factor = 0.01 / sum_initial\n seq = seq * scale_factor\n best_seq = seq.copy()\n best_value = evaluate_sequence(best_seq.tolist())\n\n # Gradually upscale the sequence length with exact length scaling\n current_len = len(best_seq)\n while current_len < target_length:\n # Double the length\n upsampling_factor = 2\n upsampled_seq = np.repeat(best_seq, upsampling_factor)\n # First scale to max constraint\n max_upsampled = np.max(upsampled_seq)\n if max_upsampled > 1000.0:\n scale_factor = 1000.0 / max_upsampled\n upsampled_seq = upsampled_seq * scale_factor\n # Now scale to sum 0.01\n sum_upsampled = np.sum(upsampled_seq)\n if sum_upsampled == 0:\n sum_upsampled = 1.0\n scale_factor = 0.01 / sum_upsampled\n upsampled_seq = upsampled_seq * scale_factor\n best_seq = upsampled_seq.copy()\n current_len = len(best_seq)\n\n # Final length adjustment to target_length\n best_seq = best_seq[:target_length]\n\n # Final optimization with improved Adam parameters and constraint maintenance\n start_time = time.time()\n best_seq = best_seq.copy()\n best_value = evaluate_sequence(best_seq.tolist())\n\n # Parameters for optimization\n max_iter = 300000 # Increased iterations for deeper exploration\n learning_rate = 0.01 # Increased learning rate for faster convergence\n decay_rate = 0.999 # Slightly less aggressive decay\n min_lr = 1e-5 # Minimum learning rate\n decay_steps = 30 # Reduced for more frequent decay\n step_counter = 0\n beta1 = 0.9\n beta2 = 0.999\n epsilon = 1e-8\n\n # Adam parameters\n m = np.zeros_like(best_seq)\n v = np.zeros_like(best_seq)\n t = 0\n\n for _ in range(max_iter):\n step_counter += 1\n # Compute gradient\n conv = np.convolve(best_seq, best_seq, mode='full')\n l2_sq, g_l2 = _simpson_l2sq(conv)\n l1, g_l1 = _l1(conv)\n linf, g_linf = _linf(conv)\n\n if l1 <= 0.0 or linf <= 0.0:\n break # Avoid division by zero\n\n denom = l1 * linf\n c_value = l2_sq / denom\n\n # Compute gradient of C with respect to conv\n g_conv = _objective_and_grad_conv(conv)[1]\n\n # Compute gradient of C with respect to h\n grad_h = _grad_h_from_conv_grad(best_seq, g_conv)\n\n # Adam update\n t += 1\n m = beta1 * m + (1 - beta1) * grad_h\n v = beta2 * v + (1 - beta2) * (grad_h ** 2)\n m_hat = m / (1 - beta1 ** t)\n v_hat = v / (1 - beta2 ** t)\n step = learning_rate * m_hat / (np.sqrt(v_hat) + epsilon)\n best_seq = best_seq + step\n\n # Ensure values are within bounds\n best_seq = np.clip(best_seq, 0.0, 1000.0)\n\n # Maintain minimum sum requirement\n sum_seq = np.sum(best_seq)\n if sum_seq < 0.01:\n scale = 0.01 / sum_seq\n best_seq = best_seq * scale\n best_seq = np.clip(best_seq, 0.0, 1000.0)\n sum_seq = np.sum(best_seq)\n if sum_seq < 0.01:\n best_seq = np.zeros_like(best_seq)\n best_seq[0] = 0.01\n\n # Track best value\n if step_counter % 50 == 0:\n current_value = evaluate_sequence(best_seq.tolist())\n if current_value > best_value:\n best_value = current_value\n best_seq = best_seq.copy()\n\n # Check time limit\n if time.time() - start_time > max_seconds - 5:\n break\n\n # Decay learning rate\n if step_counter % decay_steps == 0:\n learning_rate *= decay_rate\n\n # Final evaluation\n final_value = evaluate_sequence(best_seq.tolist())\n print(f\"Final C2 lower bound: {final_value:.6f}\")\n return best_seq.tolist()\n```",
64 "env/all/time/policy": 725.0600522314198,
65 "env/all/time/policy/min": 340.1176002025604,
66 "env/all/time/policy/max": 1030.7621109485626,
67 "env/all/time/env_step": 2303.706905388739,
68 "env/all/time/env_step/min": 0.009292364120483398,
69 "env/all/time/env_step/max": 4388.473441362381,
70 "env/all/time/reward_compute": 6.300397217273712e-07,
71 "env/all/time/reward_compute/min": 2.2351741790771484e-07,
72 "env/all/time/reward_compute/max": 1.4565885066986084e-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": 0.021334102377295494,
77 "advantage/min": -1.0,
78 "advantage/max": 5.788096904754639,
79 "time/assemble_training_data": 11.401707172393799,
80 "time/kl_vs_base": 173.6981201171875,
81 "kl_policy_base": 0.0006401613936759531,
82 "time/train": 1432.0798869132996,
83 "time/save_checkpoint": 15.319638967514038,
84 "time/total": 7030.976446866989
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
[2026-07-12T09:30:52+00:00] job=1829971 node=node-7 ngpu=3 ntrain=1 replicas=2 flash_attn=yes