Views
No views yet
ac2. 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": 550,
7 "puct/sampled_size": 8,
8 "puct/T": 17408,
9 "puct/scale_last": 0.4401506157434054,
10 "puct/buffer_value/mean": 0.9322674608783944,
11 "puct/buffer_value/std": 0.04166216278553881,
12 "puct/buffer_value/min": 0.5041471954736918,
13 "puct/buffer_value/max": 0.9442978112170972,
14 "puct/buffer_timestep/mean": 16.23818181818182,
15 "puct/buffer_timestep/std": 9.978139577170065,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 33.0,
18 "puct/buffer_construction_len/mean": 2929.1454545454544,
19 "puct/buffer_construction_len/std": 2392.2642672366946,
20 "puct/buffer_construction_len/min": 1024.0,
21 "puct/buffer_construction_len/max": 32768.0,
22 "puct/sampled_value/mean": 0.9442698008548515,
23 "puct/sampled_value/std": 1.5988436972260585e-05,
24 "puct/sampled_value/min": 0.9442511664293527,
25 "puct/sampled_value/max": 0.9442978112170972,
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": 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": 4291.703719615936,
35 "env/all/ac_tokens_per_turn": 9699.001953125,
36 "env/all/ob_tokens_per_turn": 3979.5,
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": 4965889,
41 "env/all/total_ob_tokens": 2037504,
42 "env/all/time/sampling_mean": 389.02294355770573,
43 "env/all/time/sampling_max": 554.7691810131073,
44 "env/all/time/env_step_mean": 1757.2079339195043,
45 "env/all/time/env_step_max": 3757.0440723896027,
46 "env/all/reward/mean": 0.7675218342845648,
47 "env/all/reward/max": 0.9443503155233095,
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": 0.7675218342845648,
53 "env/all/correctness": 0.8359375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.9181569606394795,
57 "env/all/raw_score/min": 0.02679886254213666,
58 "env/all/raw_score/max": 0.9443503155233095,
59 "env/all/initial_raw_score": 0.9442698008548516,
60 "env/all/initial_raw_score/min": 0.9442511664293527,
61 "env/all/initial_raw_score/max": 0.9442978112170972,
62 "env/all/msg": "Success; raw_score=0.9435805760996453",
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 # Extract the central N elements for the gradient\n gradient = g_h[N-1 : N-1 + N]\n return gradient\n\ndef construct_function():\n \"\"\"Construct a sequence to maximize the lower bound C using adaptive upscaling, more diverse\n initial sequences, refined gradient ascent with constraint maintenance, and better scaling.\"\"\"\n # Parameters\n n_start = 512\n target_length = 4096\n max_seconds = 1000\n num_initializations = 16 # Increase diversity to avoid local minima\n learning_rate = 0.02\n decay_rate = 0.999\n min_lr = 1e-6\n decay_steps = 100\n max_iter = 200000\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 initial sequences with sum of 0.01\n for _ in range(num_initializations):\n # Generate a random sequence with sum = 0.01 and values in [0, 1000]\n seq = np.random.uniform(low=0.0, high=1000.0, size=n_start)\n sum_seq = np.sum(seq)\n # Ensure sum is not too small\n while sum_seq < 0.01:\n seq = np.random.uniform(low=0.0, high=1000.0, size=n_start)\n sum_seq = np.sum(seq)\n scale = 0.01 / sum_seq\n seq = seq * scale\n initial_sequences.append(seq)\n\n # Generate structured sequence with multiple peaks\n seq = np.zeros(n_start, dtype=np.float32)\n num_peaks = random.randint(4, 8)\n idxs = np.random.choice(n_start, num_peaks, replace=False)\n for i in idxs:\n seq[i] = 0.01 / num_peaks\n initial_sequences.append(seq)\n\n # Generate sparse sequence with one high peak\n seq = np.zeros(n_start, dtype=np.float32)\n peak_idx = random.randint(0, n_start - 1)\n seq[peak_idx] = 0.01\n initial_sequences.append(seq)\n\n # Generate 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 sum_seq = np.sum(seq)\n scale = 0.01 / sum_seq\n seq = seq * scale\n initial_sequences.append(seq)\n\n # Generate exponential 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 sum_seq = np.sum(seq)\n scale = 0.01 / sum_seq\n seq = seq * scale\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 a random sequence\n seq = np.random.uniform(low=0.0, high=1000.0, size=n_start)\n sum_seq = np.sum(seq)\n scale = 0.01 / sum_seq\n best_seq = seq * scale\n best_value = evaluate_sequence(best_seq.tolist())\n\n # Gradually upscale the sequence length\n current_len = len(best_seq)\n while current_len < target_length:\n upsampling_factor = max(2, (target_length + current_len - 1) // current_len)\n upsampled_seq = np.repeat(best_seq, upsampling_factor)\n best_seq = upsampled_seq[:target_length]\n current_len = len(best_seq)\n\n # Final optimization with adaptive Adam optimization\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 m = np.zeros_like(best_seq)\n v = np.zeros_like(best_seq)\n t = 0\n beta1 = 0.9\n beta2 = 0.999\n epsilon = 1e-8\n learning_rate = 0.02\n max_iter = 200000\n\n for _ in range(max_iter):\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 # Clip parameters to avoid clipping (no sum constraint)\n best_seq = np.clip(best_seq, 0.0, 1000.0)\n\n # Track best value\n if _ % 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 _ % 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": 389.02294355770573,
65 "env/all/time/policy/min": 133.58788681030273,
66 "env/all/time/policy/max": 554.7691810131073,
67 "env/all/time/env_step": 1757.2079339195043,
68 "env/all/time/env_step/min": 0.008790016174316406,
69 "env/all/time/env_step/max": 3757.0440723896027,
70 "env/all/time/reward_compute": 2.1560117602348328e-07,
71 "env/all/time/reward_compute/min": 1.7881393432617188e-07,
72 "env/all/time/reward_compute/max": 3.3527612686157227e-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": 0.01922135055065155,
77 "advantage/min": -1.0,
78 "advantage/max": 4.184759140014648,
79 "time/assemble_training_data": 10.15076231956482,
80 "time/kl_vs_base": 104.81732225418091,
81 "kl_policy_base": 0.0006706287967972457,
82 "time/train": 677.2204778194427,
83 "time/save_checkpoint": 31.884238719940186,
84 "time/total": 5120.428997993469
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