Views
No views yet
ac2. Checkpoint saved
after training step 32 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 32,
3 "progress/batch": 32,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.66,
6 "puct/buffer_size": 517,
7 "puct/sampled_size": 8,
8 "puct/T": 16384,
9 "puct/scale_last": 0.28145519259097795,
10 "puct/buffer_value/mean": 0.9348026821190942,
11 "puct/buffer_value/std": 0.04034913444285277,
12 "puct/buffer_value/min": 0.6666666666666636,
13 "puct/buffer_value/max": 0.9481218592576446,
14 "puct/buffer_timestep/mean": 15.317214700193423,
15 "puct/buffer_timestep/std": 9.364112818443314,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 31.0,
18 "puct/buffer_construction_len/mean": 3826.2727272727275,
19 "puct/buffer_construction_len/std": 1778.804607269239,
20 "puct/buffer_construction_len/min": 1024.0,
21 "puct/buffer_construction_len/max": 32768.0,
22 "puct/sampled_value/mean": 0.9481126444466885,
23 "puct/sampled_value/std": 6.272791709947608e-06,
24 "puct/sampled_value/min": 0.9480995672502311,
25 "puct/sampled_value/max": 0.9481218592576446,
26 "puct/sampled_timestep/mean": 31.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 31.0,
29 "puct/sampled_timestep/max": 31.0,
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": 5714.066039085388,
35 "env/all/ac_tokens_per_turn": 8333.5234375,
36 "env/all/ob_tokens_per_turn": 4148.75,
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": 4266764,
41 "env/all/total_ob_tokens": 2124160,
42 "env/all/time/sampling_mean": 564.309276740998,
43 "env/all/time/sampling_max": 779.162454366684,
44 "env/all/time/env_step_mean": 2463.7420410686173,
45 "env/all/time/env_step_max": 4952.540219545364,
46 "env/all/reward/mean": 0.3764312541727313,
47 "env/all/reward/max": 0.9481221327632995,
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": 0.3764312541727313,
53 "env/all/correctness": 0.40234375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.9355961268759148,
57 "env/all/raw_score/min": 0.663161937802744,
58 "env/all/raw_score/max": 0.9481221327632995,
59 "env/all/initial_raw_score": 0.9481126444466885,
60 "env/all/initial_raw_score/min": 0.9480995672502311,
61 "env/all/initial_raw_score/max": 0.9481218592576446,
62 "env/all/msg": "Success; raw_score=0.948121888360036",
63 "env/all/parsed_code": "```python\nimport numpy as np\nimport time\nfrom typing import List, Tuple\nimport random\n\ndef _simpson_l2sq(conv: np.ndarray) -> Tuple[float, np.ndarray]:\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 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 if conv.size == 0:\n return 0.0, np.zeros_like(conv)\n m = float(np.max(conv))\n if m <= 0.0:\n return m, np.zeros_like(conv)\n mask = conv == m\n count = int(mask.sum())\n if count == 0:\n return m, np.zeros_like(conv)\n grad = mask.astype(conv.dtype) / count\n return m, grad\n\ndef _objective_and_grad_conv(conv: np.ndarray) -> Tuple[float, np.ndarray]:\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 h_rev = h[::-1]\n g_h = np.convolve(g_conv, h_rev, mode=\"valid\")\n return 2.0 * g_h\n\ndef _upsample_1d(h: np.ndarray) -> np.ndarray:\n n = h.shape[0]\n x_old = np.linspace(-0.5, 0.5, n)\n x_new = np.linspace(-0.5, 0.5, 2 * n)\n return np.interp(x_new, x_old, h)\n\ndef construct_function():\n \"\"\"\n This function constructs a sequence of non-negative heights to maximize the evaluation function.\n It uses a hybrid approach with enhanced exploration, larger initial noise, better learning rate decay,\n and dynamic restarts to escape local optima.\n \"\"\"\n np.random.seed(42)\n height_sequence_1 = globals().get(\"height_sequence_1\", None)\n target_length = 4096\n initial_n = 8\n\n # Generate diverse initial sequences\n if height_sequence_1 is not None:\n initial_h = np.array(height_sequence_1, dtype=np.float32)\n initial_n = min(len(initial_h), target_length)\n x_old = np.linspace(-0.5, 0.5, len(initial_h))\n x_new = np.linspace(-0.5, 0.5, initial_n)\n h = np.interp(x_new, x_old, initial_h)\n else:\n candidates = []\n # Gaussian peaks with varying number of peaks\n for num_peaks in range(3, 8):\n x = np.linspace(-0.5, 0.5, target_length)\n peaks_positions = np.linspace(-0.3, 0.3, num_peaks)\n seq = np.zeros(target_length)\n for pos in peaks_positions:\n seq += 0.05 * np.exp(-((x - pos) / 0.15)**2)\n seq = np.clip(seq / np.sum(seq) * 0.01, 0.0, 1000.0)\n candidates.append(seq)\n # Random sequences\n for _ in range(5):\n seq = np.random.rand(target_length) * 0.01\n seq = np.clip(seq / np.sum(seq) * 0.01, 0.0, 1000.0)\n candidates.append(seq)\n # Sinusoidal patterns\n seq_sw = 0.05 * np.sin(2 * np.pi * np.linspace(-0.5, 0.5, target_length) * 3 + np.random.rand(10))\n seq_sw = np.clip(seq_sw / np.sum(seq_sw) * 0.01, 0.0, 1000.0)\n candidates.append(seq_sw)\n # Exponential decay with multiple peaks\n seq_decay = 0.05 * np.exp(-np.abs(np.linspace(-0.5, 0.5, target_length)) * 5)\n seq_decay = np.clip(seq_decay / np.sum(seq_decay) * 0.01, 0.0, 1000.0)\n candidates.append(seq_decay)\n # Sharp central peak\n seq_peak = np.zeros(target_length)\n seq_peak[target_length // 2] = 0.01\n seq_peak = np.clip(seq_peak, 0.0, 1000.0)\n candidates.append(seq_peak)\n # Select the best candidate\n best_c = -1.0\n best_h = candidates[0]\n for h_candidate in candidates:\n try:\n c = evaluate_sequence(h_candidate.tolist())\n if c > best_c:\n best_c = c\n best_h = h_candidate\n except:\n pass\n h = best_h\n\n # Parameters for enhanced exploration and refinement\n learning_rate_initial = 0.3\n noise_scale_initial = 30.0\n noise_decay = 0.98\n learning_rate_decay = 0.99\n max_steps = 750000\n upscale_steps = 200000\n refine_steps = 500000\n restart_threshold = 1000 # Steps without improvement to trigger restart\n\n start_time = time.time()\n\n # Multi-scale optimization with enhanced exploration\n current_length = initial_n\n scale_factor = 2\n\n for _ in range(5):\n if current_length < target_length:\n h = _upsample_1d(h)\n current_length *= 2\n h_sum = np.sum(h)\n if h_sum < 0.01:\n h = np.clip(h + (0.01 - h_sum) / h.shape[0], 0.0, 1000.0)\n\n # Initial optimization with dynamic noise and learning rate\n h_opt = np.copy(h)\n h_best = np.copy(h)\n best_c = -1.0\n no_improvement_steps = 0\n\n for step in range(upscale_steps):\n clipped_h = np.clip(h_opt, 0.0, None)\n conv = np.convolve(clipped_h, clipped_h, mode='full')\n obj_val, grad_conv = _objective_and_grad_conv(conv)\n grad_h = _grad_h_from_conv_grad(clipped_h, grad_conv)\n\n noise_scale = noise_scale_initial * (noise_decay ** step)\n learning_rate_current = learning_rate_initial * (learning_rate_decay ** step)\n\n noise = noise_scale * np.random.normal(size=h_opt.shape)\n h_opt = np.clip(h_opt + learning_rate_current * grad_h + noise, 0.0, 1000.0)\n\n h_sum_current = np.sum(h_opt)\n if h_sum_current < 0.01:\n scale_factor = 0.01 / h_sum_current\n h_opt = np.clip(h_opt * scale_factor, 0.0, 1000.0)\n\n if step % 400 == 0:\n try:\n current_c = evaluate_sequence(h_opt.tolist())\n if current_c > best_c:\n best_c = current_c\n h_best = h_opt.copy()\n no_improvement_steps = 0\n else:\n no_improvement_steps += 1\n except:\n pass\n\n if no_improvement_steps >= restart_threshold:\n # Trigger restart\n print(\"Restarting optimization due to lack of improvement.\")\n h_opt = np.random.rand(h.shape[0]) * 0.01\n h_opt = np.clip(h_opt / np.sum(h_opt) * 0.01, 0.0, 1000.0)\n no_improvement_steps = 0\n\n remaining_time = 1000 - (time.time() - start_time)\n if remaining_time < 3:\n print(f\"Time remaining: {remaining_time} seconds. Performing final refinement.\")\n break\n h = h_best\n else:\n break\n\n # Final refined optimization with aggressive learning rate and adaptive steps\n h_refined = np.copy(h)\n for step in range(refine_steps):\n clipped_h = np.clip(h_refined, 0.0, None)\n conv = np.convolve(clipped_h, clipped_h, mode='full')\n obj_val, grad_conv = _objective_and_grad_conv(conv)\n grad_h = _grad_h_from_conv_grad(clipped_h, grad_conv)\n\n learning_rate_current = learning_rate_initial * (learning_rate_decay ** step)\n h_refined = np.clip(h_refined + learning_rate_current * grad_h, 0.0, 1000.0)\n\n remaining_time = 1000 - (time.time() - start_time)\n if remaining_time < 3:\n print(f\"Time remaining: {remaining_time} seconds. Final step.\")\n break\n\n h_final = np.clip(h_refined, 0.0, 1000.0)\n heights = h_final.tolist()\n r_value = evaluate_sequence(heights)\n print(f\"Final C2 lower bound: {r_value}\")\n return heights\n```",
64 "env/all/time/policy": 564.309276740998,
65 "env/all/time/policy/min": 296.40004420280457,
66 "env/all/time/policy/max": 779.162454366684,
67 "env/all/time/env_step": 2463.7420410686173,
68 "env/all/time/env_step/min": 0.008813619613647461,
69 "env/all/time/env_step/max": 4952.540219545364,
70 "env/all/time/reward_compute": 4.591420292854309e-07,
71 "env/all/time/reward_compute/min": 2.1606683731079102e-07,
72 "env/all/time/reward_compute/max": 8.23289155960083e-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": 0.023418404161930084,
77 "advantage/min": -1.0,
78 "advantage/max": 4.715853691101074,
79 "time/assemble_training_data": 6.097881078720093,
80 "time/kl_vs_base": 151.73404002189636,
81 "kl_policy_base": 0.0007284191669896245,
82 "time/train": 1207.409583568573,
83 "time/save_checkpoint": 10.284076452255249,
84 "time/total": 7094.838932514191
85}[2026-07-09T06:42:12+00:00] job=1812634 node=node-30 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:26:33+00:00] job=1812736 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:27:32+00:00] job=1813133 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T16:26:04+00:00] job=1813134 node=node-4 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-11T01:56:04+00:00] job=1821290 node=node-4 ngpu=3 ntrain=1 replicas=2 flash_attn=yes