Views
No views yet
ac2. Checkpoint saved
after training step 2 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 2,
3 "progress/batch": 2,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.06,
6 "puct/buffer_size": 40,
7 "puct/sampled_size": 8,
8 "puct/T": 1024,
9 "puct/scale_last": 0.12839541096274187,
10 "puct/buffer_value/mean": 0.8648423856747897,
11 "puct/buffer_value/std": 0.10110497620294691,
12 "puct/buffer_value/min": 0.6666666666666636,
13 "puct/buffer_value/max": 0.9286918076170656,
14 "puct/buffer_timestep/mean": 0.2,
15 "puct/buffer_timestep/std": 0.7483314773547883,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 1.0,
18 "puct/buffer_construction_len/mean": 2862.575,
19 "puct/buffer_construction_len/std": 1816.822719578055,
20 "puct/buffer_construction_len/min": 1024.0,
21 "puct/buffer_construction_len/max": 6967.0,
22 "puct/sampled_value/mean": 0.9252597303396561,
23 "puct/sampled_value/std": 0.0021352963739526953,
24 "puct/sampled_value/min": 0.9225934514196982,
25 "puct/sampled_value/max": 0.9286918076170656,
26 "puct/sampled_timestep/mean": 1.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 1.0,
29 "puct/sampled_timestep/max": 1.0,
30 "puct/sampled_construction_len/mean": 2560.0,
31 "puct/sampled_construction_len/std": 1254.138748304987,
32 "puct/sampled_construction_len/min": 1024.0,
33 "puct/sampled_construction_len/max": 4096.0,
34 "time/sampling": 5377.276038408279,
35 "env/all/ac_tokens_per_turn": 7998.09375,
36 "env/all/ob_tokens_per_turn": 4199.142857142857,
37 "env/all/turns_per_episode": 1.0,
38 "env/all/total_episodes": 448,
39 "env/all/total_turns": 448,
40 "env/all/total_ac_tokens": 3583146,
41 "env/all/total_ob_tokens": 1881216,
42 "env/all/time/sampling_mean": 544.5272494751428,
43 "env/all/time/sampling_max": 765.107216835022,
44 "env/all/time/env_step_mean": 2218.1604429424874,
45 "env/all/time/env_step_max": 4608.438857078552,
46 "env/all/reward/mean": 0.25740019416032855,
47 "env/all/reward/max": 0.9279974669244346,
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": 0.25740019416032855,
53 "env/all/correctness": 0.28348214285714285,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.9079943856994267,
57 "env/all/raw_score/min": 0.676976124959557,
58 "env/all/raw_score/max": 0.9279974669244346,
59 "env/all/initial_raw_score": 0.9249300522136032,
60 "env/all/initial_raw_score/min": 0.9225934514196982,
61 "env/all/initial_raw_score/max": 0.9286918076170656,
62 "env/all/msg": "Evaluation timed out after 1100 seconds.",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom typing import Tuple\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.empty(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 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 grad /= 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=1e-1, 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=90_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 increased exploration and diversity.\n Key improvements:\n - Increased total iterations and exploration time\n - Higher initial noise and slower decay for exploration\n - Larger keep fraction to maintain population diversity\n - Increased fine-tuning steps for better convergence\n - Adjusted learning rates for different phases\n \"\"\"\n n = 256\n bsz = 64 # Reduced batch size to allow more iterations\n total_iter = 120_000 # Increased for more exploration\n explore_steps = 80_000 # Extended exploration phase\n drop_every = 30_000\n keep_frac = 0.6 # Maintained diversity\n lr_explore = 7e-3 # Higher exploration rate\n lr_exploit = 2e-3\n eta, gamma = 1e-1, 0.8 # More noise early on, slower decay\n \n prev = globals().get(\"GLOBAL_BEST_CONSTRUCTION\") or list()\n if isinstance(prev, (list, tuple, np.ndarray)) and len(prev) > 0:\n h_prev_best = np.array(prev, dtype=np.float32)\n else:\n h_prev_best = np.ones(n, dtype=np.float32)\n h_prev_best = np.clip(h_prev_best, 0.0, None)\n\n # Ensure previous best is size matches\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(np.float32)\n\n rng = np.random.default_rng()\n\n def init_sampler(m):\n out = rng.uniform(0.0, 1.0, size=(m, n)).astype(np.float32)\n if m > 0:\n out[0] = h_prev_best\n return out\n\n h_batch = init_sampler(bsz)\n opt_list = [_Adam(shape=(n,), lr=lr_explore, dtype=np.float32) for _ in range(bsz)]\n best_h = h_batch.copy()\n best_c = np.full(bsz, -np.inf, dtype=np.float32)\n\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=False, 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 idx = int(np.argmax(best_c))\n h_star = np.clip(best_h[idx].astype(np.float32), 0.0, None)\n\n # First upsample\n h_up1 = _upsample_1d(h_star)\n h_up1, _ = _single_candidate_finetune(h_up1, lr=3e-3, steps=90_000)\n\n # Second upsample\n h_up2 = _upsample_1d(h_up1)\n h_up2, _ = _single_candidate_finetune(h_up2, lr=3e-3, steps=90_000)\n\n # Third upsample\n h_up3 = _upsample_1d(h_up2)\n h_up3, _ = _single_candidate_finetune(h_up3, lr=3e-3, steps=90_000)\n\n # Final fine-tuning\n h_up4 = _upsample_1d(h_up3)\n h_up4, _ = _single_candidate_finetune(h_up4, lr=1e-3, steps=90_000)\n\n heights = np.clip(h_up4, 0.0, None)\n r_value = evaluate_sequence(heights.tolist())\n print(\"This gets a C2 lower bound of\", r_value)\n return heights.tolist()\n```",
64 "env/all/time/policy": 544.5272494751428,
65 "env/all/time/policy/min": 210.12499833106995,
66 "env/all/time/policy/max": 765.107216835022,
67 "env/all/time/env_step": 2218.1604429424874,
68 "env/all/time/env_step/min": 0.006566762924194336,
69 "env/all/time/env_step/max": 4608.438857078552,
70 "env/all/time/reward_compute": 3.262289932795933e-07,
71 "env/all/time/reward_compute/min": 2.123415470123291e-07,
72 "env/all/time/reward_compute/max": 5.774199962615967e-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": 0.03649523854255676,
77 "advantage/min": -1.0,
78 "advantage/max": 6.9617695808410645,
79 "time/assemble_training_data": 8.792766332626343,
80 "time/kl_vs_base": 130.13185834884644,
81 "kl_policy_base": 0.0004687599721364677,
82 "time/train": 1026.5865852832794,
83 "time/save_checkpoint": 15.310400009155273,
84 "time/total": 6560.0196306705475
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