Views
No views yet
ac2. Checkpoint saved
after training step 6 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 6,
3 "progress/batch": 6,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.14,
6 "puct/buffer_size": 101,
7 "puct/sampled_size": 8,
8 "puct/T": 3072,
9 "puct/scale_last": 0.18502301457897785,
10 "puct/buffer_value/mean": 0.8977844508231578,
11 "puct/buffer_value/std": 0.0732405566428329,
12 "puct/buffer_value/min": 0.6666666666666636,
13 "puct/buffer_value/max": 0.941478874818969,
14 "puct/buffer_timestep/mean": 2.207920792079208,
15 "puct/buffer_timestep/std": 1.8948436111082592,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 5.0,
18 "puct/buffer_construction_len/mean": 3212.108910891089,
19 "puct/buffer_construction_len/std": 3354.2366763852133,
20 "puct/buffer_construction_len/min": 1024.0,
21 "puct/buffer_construction_len/max": 32768.0,
22 "puct/sampled_value/mean": 0.9337830138167975,
23 "puct/sampled_value/std": 0.003974998063161984,
24 "puct/sampled_value/min": 0.9300009684046523,
25 "puct/sampled_value/max": 0.941478874818969,
26 "puct/sampled_timestep/mean": 5.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 5.0,
29 "puct/sampled_timestep/max": 5.0,
30 "puct/sampled_construction_len/mean": 2816.0,
31 "puct/sampled_construction_len/std": 1330.2150202128978,
32 "puct/sampled_construction_len/min": 1024.0,
33 "puct/sampled_construction_len/max": 4096.0,
34 "time/sampling": 5509.390102148056,
35 "env/all/ac_tokens_per_turn": 8109.654296875,
36 "env/all/ob_tokens_per_turn": 4236.0,
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": 4152143,
41 "env/all/total_ob_tokens": 2168832,
42 "env/all/time/sampling_mean": 301.92568093398586,
43 "env/all/time/sampling_max": 487.16717171669006,
44 "env/all/time/env_step_mean": 2527.4041273370385,
45 "env/all/time/env_step_max": 5116.264766931534,
46 "env/all/reward/mean": 0.3519246175772517,
47 "env/all/reward/max": 0.9433691051292941,
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": 0.3519246175772517,
53 "env/all/correctness": 0.392578125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.8964447970127009,
57 "env/all/raw_score/min": 0.04441102939792563,
58 "env/all/raw_score/max": 0.9433691051292941,
59 "env/all/initial_raw_score": 0.9337830138167974,
60 "env/all/initial_raw_score/min": 0.9300009684046523,
61 "env/all/initial_raw_score/max": 0.941478874818969,
62 "env/all/msg": "RuntimeError: Program execution failed: ImportError: cannot import name 'np' from 'typing' (/usr/lib/python3.12/typing.py)",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom typing import List, Tuple, np\nimport time\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\nclass _Adam:\n def __init__(self, shape, lr=3e-2, beta1=0.9, beta2=0.999, eps=1e-8, dtype=np.float32):\n self.m = np.zeros(shape, dtype=dtype)\n self.v = np.zeros(shape, dtype=dtype)\n self.t = 0\n self.lr = lr\n self.b1 = beta1\n self.b2 = beta2\n self.eps = eps\n\n def step(self, params, grad):\n self.t += 1\n self.m = self.b1 * self.m + (1 - self.b1) * grad\n self.v = self.b2 * self.v + (1 - self.b2) * (grad * grad)\n m_hat = self.m / (1 - self.b1 ** self.t)\n v_hat = self.v / (1 - self.b2 ** self.t)\n return params + self.lr * m_hat / (np.sqrt(v_hat) + self.eps)\n\ndef _batch_objective(h_batch: np.ndarray) -> Tuple[np.ndarray, list[np.ndarray]]:\n bsz = h_batch.shape[0]\n c_vals = np.zeros(bsz, dtype=np.float32)\n conv_grads = [None] * bsz\n for b in range(bsz):\n h = np.clip(h_batch[b], 0.0, None)\n conv = np.convolve(h, h, mode=\"full\")\n c_val, g_conv = _objective_and_grad_conv(conv)\n c_vals[b] = c_val\n conv_grads[b] = g_conv\n return c_vals, conv_grads\n\ndef _phase_update(h_batch, opt_list, lr, add_noise=False, t=0, eta=0.5, gamma=0.8):\n bsz = h_batch.shape[0]\n c_vals, conv_grads = _batch_objective(h_batch)\n grads = np.zeros_like(h_batch, dtype=h_batch.dtype)\n for b in range(bsz):\n clipped = np.clip(h_batch[b], 0.0, None)\n grads[b] = _grad_h_from_conv_grad(clipped, conv_grads[b])\n if add_noise:\n sigma = eta / ((t + 1) ** gamma)\n grads = grads + sigma * np.random.normal(size=grads.shape).astype(grads.dtype)\n for b in range(bsz):\n opt = opt_list[b]\n opt.lr = lr\n h_new = opt.step(h_batch[b], grads[b].astype(h_batch.dtype))\n h_batch[b] = np.clip(h_new, 0.0, None)\n return h_batch, c_vals\n\ndef _elitist_respawn(h_batch, c_vals, keep_frac, init_sampler, opt_list):\n bsz = h_batch.shape[0]\n keep_n = max(1, int(bsz * keep_frac))\n idx = np.argsort(c_vals)[-keep_n:]\n survivors = h_batch[idx].copy()\n fresh = init_sampler(bsz - keep_n)\n new_batch = np.concatenate([survivors, fresh], axis=0)\n new_opts = []\n for i in range(keep_n):\n new_opts.append(opt_list[idx[i]])\n for _ in range(bsz - keep_n):\n new_opts.append(_Adam(shape=h_batch.shape[1:], lr=opt_list[0].lr, dtype=h_batch.dtype))\n return new_batch, new_opts\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 _single_candidate_finetune(h0: np.ndarray, lr=3e-3, steps=300_000) -> Tuple[np.ndarray, float]:\n h = h0.astype(np.float32).copy()\n opt = _Adam(h.shape, lr=lr, dtype=h.dtype)\n last_c = 0.0\n for _ in range(steps):\n h_clip = np.clip(h, 0.0, None)\n conv = np.convolve(h_clip, h_clip, mode=\"full\")\n c_val, g_conv = _objective_and_grad_conv(conv)\n g_h = _grad_h_from_conv_grad(h_clip, g_conv)\n h = np.clip(opt.step(h, g_h.astype(h.dtype)), 0.0, None)\n last_c = c_val\n return h, float(last_c)\n\ndef construct_function():\n \"\"\"\n Optimized gradient-based search with diversified initialization, adaptive learning, and strategic fine-tuning.\n Integrates structured pattern-based initialization, dynamic noise scaling, and enhanced fine-tuning to maximize the lower bound.\n \"\"\"\n np.random.seed(42)\n \n height_sequence_1 = globals().get(\"height_sequence_1\", None)\n n = 512\n bsz = 64\n total_iter = 150000 # Increased for more exploration\n explore_steps = 80000 # Extended exploration phase\n drop_every = 3000\n keep_frac = 0.6 # Maintain moderate diversity\n lr_explore = 3e-2 # Increased exploration learning rate\n lr_exploit = 1e-4\n eta, gamma = 0.5, 0.8 # More aggressive noise decay\n dtype = np.float32\n\n prev = globals().get(\"GLOBAL_BEST_CONSTRUCTION\", [])\n if isinstance(prev, (list, tuple, np.ndarray)) and len(prev) > 0:\n h_prev_best = np.array(prev, dtype=dtype)\n else:\n h_prev_best = np.ones(n, dtype=dtype)\n h_prev_best = np.clip(h_prev_best, 0.0, None)\n\n if h_prev_best.shape[0] != n:\n x_old = np.linspace(-0.5, 0.5, h_prev_best.shape[0])\n x_new = np.linspace(-0.5, 0.5, n)\n h_prev_best = np.interp(x_new, x_old, h_prev_best).astype(dtype)\n\n def init_sampler(m):\n \"\"\"Generate structured and random initial candidates with varied patterns\"\"\"\n out = np.zeros((m, n), dtype=dtype)\n # Scale-based variations if sequence is available\n if height_sequence_1 is not None:\n h_init = np.array(height_sequence_1, dtype=dtype)\n h_init = np.interp(np.linspace(-0.5, 0.5, n), np.linspace(-0.5, 0.5, len(h_init)), h_init)\n for i in range(m):\n out[i] = h_init * np.random.uniform(0.6, 1.4)\n # Random base candidates with structured patterns\n for i in range(m):\n if height_sequence_1 is None or i >= m // 2:\n # Random spikes or cosine-like patterns\n base = np.zeros(n, dtype=dtype)\n if np.random.rand() < 0.6:\n # Random spikes\n num_peaks = np.random.randint(2, 5)\n peak_positions = np.random.choice(n, size=num_peaks, replace=False)\n peak_heights = np.random.uniform(0.5, 2.0, size=num_peaks)\n for pos, height in zip(peak_positions, peak_heights):\n base[pos] = height\n else:\n # Smooth cosine-like pattern\n freq = np.random.uniform(0.1, 0.9)\n phase = np.random.uniform(0, 2 * np.pi)\n base = 0.5 * (1 + np.cos(2 * np.pi * freq * np.linspace(0, 1, n) + phase))\n else:\n # Gaussian-like pattern\n x = np.linspace(-0.5, 0.5, n)\n base = np.exp(-0.5 * (x / 0.25)**2)\n out[i] = np.abs(base) * np.random.uniform(1.0, 1.5)\n return out\n\n h_batch = init_sampler(bsz)\n opt_list = [_Adam(shape=(n,), lr=lr_explore, dtype=dtype) for _ in range(bsz)]\n best_h = h_batch.copy()\n best_c = np.full(bsz, -np.inf, dtype=dtype)\n\n start_time = time.time()\n for t in range(total_iter):\n if t < explore_steps:\n h_batch, c_vals = _phase_update(\n h_batch, opt_list, lr=lr_explore, add_noise=True, t=t, eta=eta, gamma=gamma\n )\n else:\n h_batch, c_vals = _phase_update(\n h_batch, opt_list, lr=lr_exploit, add_noise=True, t=t, eta=eta, gamma=gamma\n )\n\n improved = c_vals > best_c\n best_c = np.where(improved, c_vals, best_c)\n best_h[improved] = h_batch[improved]\n\n if (t + 1) % drop_every == 0:\n h_batch, opt_list = _elitist_respawn(\n h_batch, c_vals, keep_frac=keep_frac, init_sampler=init_sampler, opt_list=opt_list\n )\n\n if time.time() - start_time > 800:\n print(\"Time remaining: 200 seconds. Refining best candidate.\")\n break\n\n idx = int(np.argmax(best_c))\n h_star = np.clip(best_h[idx].astype(np.float32), 0.0, None)\n\n h_up1 = _upsample_1d(h_star)\n h_up1, _ = _single_candidate_finetune(h_up1, lr=3e-3, steps=300_000) # Increased fine-tuning steps\n\n h_up2 = _upsample_1d(h_up1)\n h_up2, _ = _single_candidate_finetune(h_up2, lr=3e-3, steps=300_000)\n\n heights = np.clip(h_up2, 0.0, None)\n r_value = evaluate_sequence(heights.tolist())\n print(f\"Final C2 lower bound: {r_value}\")\n return heights.tolist()\n```",
64 "env/all/time/policy": 301.92568093398586,
65 "env/all/time/policy/min": 87.06767344474792,
66 "env/all/time/policy/max": 487.16717171669006,
67 "env/all/time/env_step": 2527.4041273370385,
68 "env/all/time/env_step/min": 0.003964662551879883,
69 "env/all/time/env_step/max": 5116.264766931534,
70 "env/all/time/reward_compute": 4.4191256165504456e-07,
71 "env/all/time/reward_compute/min": 1.862645149230957e-07,
72 "env/all/time/reward_compute/max": 1.735985279083252e-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": 0.025444436818361282,
77 "advantage/min": -1.0,
78 "advantage/max": 5.364130020141602,
79 "time/assemble_training_data": 9.058156728744507,
80 "time/kl_vs_base": 95.32297158241272,
81 "kl_policy_base": 0.0005367703270167112,
82 "time/train": 599.035306930542,
83 "time/save_checkpoint": 9.268730640411377,
84 "time/total": 6224.5017602443695
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