Views
No views yet
ac2. Checkpoint saved
after training step 37 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 37,
3 "progress/batch": 37,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.76,
6 "puct/buffer_size": 598,
7 "puct/sampled_size": 8,
8 "puct/T": 18944,
9 "puct/scale_last": 0.4403347392430005,
10 "puct/buffer_value/mean": 0.9332394155202812,
11 "puct/buffer_value/std": 0.04009036756794236,
12 "puct/buffer_value/min": 0.5041471954736918,
13 "puct/buffer_value/max": 0.9444819347166923,
14 "puct/buffer_timestep/mean": 17.744147157190636,
15 "puct/buffer_timestep/std": 10.844897272353943,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 36.0,
18 "puct/buffer_construction_len/mean": 3022.8060200668897,
19 "puct/buffer_construction_len/std": 2316.0480986256966,
20 "puct/buffer_construction_len/min": 1024.0,
21 "puct/buffer_construction_len/max": 32768.0,
22 "puct/sampled_value/mean": 0.9444723802052943,
23 "puct/sampled_value/std": 4.7733068322158374e-06,
24 "puct/sampled_value/min": 0.9444658242507101,
25 "puct/sampled_value/max": 0.9444819347166923,
26 "puct/sampled_timestep/mean": 36.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 36.0,
29 "puct/sampled_timestep/max": 36.0,
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": 4471.62353682518,
35 "env/all/ac_tokens_per_turn": 9484.693359375,
36 "env/all/ob_tokens_per_turn": 4013.75,
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": 4856163,
41 "env/all/total_ob_tokens": 2055040,
42 "env/all/time/sampling_mean": 382.11983221769333,
43 "env/all/time/sampling_max": 483.67924094200134,
44 "env/all/time/env_step_mean": 1959.4754147455096,
45 "env/all/time/env_step_max": 3976.912698030472,
46 "env/all/reward/mean": 0.7715394025503886,
47 "env/all/reward/max": 0.9445546681450108,
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": 0.7715394025503886,
53 "env/all/correctness": 0.833984375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.9251245295217774,
57 "env/all/raw_score/min": 0.02853562361565772,
58 "env/all/raw_score/max": 0.9445546681450108,
59 "env/all/initial_raw_score": 0.9444723802052941,
60 "env/all/initial_raw_score/min": 0.9444658242507101,
61 "env/all/initial_raw_score/max": 0.9444819347166923,
62 "env/all/msg": "Success; raw_score=0.8065860993743036",
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 \"\"\"Construct a sequence to maximize the lower bound C using adaptive upscaling, better\n initial sequences, and optimized gradient ascent with constraint maintenance.\"\"\"\n # Parameters\n n_start = 128\n target_length = 4096\n max_seconds = 1000\n num_initializations = 48 # Increased diversity in initial sequences\n\n # Normalize initial sequences sum to exactly 0.01\n def generate_normalized_sequence(type='uniform'):\n if type == 'uniform':\n seq = np.full(n_start, 0.01 / n_start, dtype=np.float32)\n elif type == 'sparse':\n seq = np.zeros(n_start, dtype=np.float32)\n num_spikes = int(np.sqrt(n_start))\n selected = np.random.choice(range(n_start), num_spikes, replace=False)\n for idx in selected:\n seq[idx] = 0.01 / num_spikes\n elif type == 'gaussian':\n seq = np.zeros(n_start, dtype=np.float32)\n center = n_start // 2\n std = n_start // 6\n scale = 0.01 / np.sqrt(2 * np.pi * std**2)\n for i in range(n_start):\n seq[i] = np.exp(-((i - center) ** 2) / (2 * std**2)) * scale\n else:\n seq = np.zeros(n_start, dtype=np.float32)\n seq[0] = 0.01\n sum_seq = np.sum(seq)\n if np.abs(sum_seq - 0.01) > 1e-8:\n seq *= 0.01 / sum_seq\n return seq\n\n # Generate diverse initial sequences with sum of 0.01\n initial_sequences = []\n for _ in range(num_initializations):\n initial_sequences.append(generate_normalized_sequence('uniform'))\n initial_sequences.append(generate_normalized_sequence('sparse'))\n initial_sequences.append(generate_normalized_sequence('gaussian'))\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 uniform sequence\n seq = np.full(n_start, 0.01 / n_start, dtype=np.float32)\n best_seq = seq.copy()\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 # Ensure sum is exactly 0.01\n upsampled_sum = np.sum(upsampled_seq)\n scale_factor = 0.01 / upsampled_sum\n best_seq = upsampled_seq * scale_factor\n best_seq = best_seq[:target_length]\n current_len = len(best_seq)\n\n # Final optimization with improved Adam optimization parameters\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 = 200000 # Increased iterations for deeper exploration\n learning_rate = 0.01 # Reduced learning rate for stability\n decay_rate = 0.995 # Slower decay for more detailed exploration\n min_lr = 1e-5 # Minimum learning rate\n decay_steps = 100 # Reduced for more frequent decay\n step_counter = 0\n beta1 = 0.95\n beta2 = 0.9995\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": 382.11983221769333,
65 "env/all/time/policy/min": 187.81037163734436,
66 "env/all/time/policy/max": 483.67924094200134,
67 "env/all/time/env_step": 1959.4754147455096,
68 "env/all/time/env_step/min": 0.007124662399291992,
69 "env/all/time/env_step/max": 3976.912698030472,
70 "env/all/time/reward_compute": 2.183951437473297e-07,
71 "env/all/time/reward_compute/min": 1.6763806343078613e-07,
72 "env/all/time/reward_compute/max": 3.501772880554199e-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": 0.02100566402077675,
77 "advantage/min": -1.0,
78 "advantage/max": 4.136505603790283,
79 "time/assemble_training_data": 10.434127569198608,
80 "time/kl_vs_base": 111.8506441116333,
81 "kl_policy_base": 0.0006657962803728878,
82 "time/train": 664.9139688014984,
83 "time/save_checkpoint": 20.595086097717285,
84 "time/total": 5284.145259857178
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