Views
No views yet
ac1. Checkpoint saved
after training step 3 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 3,
3 "progress/batch": 3,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.08,
6 "puct/buffer_size": 56,
7 "puct/sampled_size": 8,
8 "puct/T": 1536,
9 "puct/scale_last": 0.39621565537633097,
10 "puct/buffer_value/mean": -1.617995778307536,
11 "puct/buffer_value/std": 0.18211313925419478,
12 "puct/buffer_value/min": -2.000000000000008,
13 "puct/buffer_value/max": -1.5094073854177394,
14 "puct/buffer_timestep/mean": 0.7142857142857143,
15 "puct/buffer_timestep/std": 1.0301575072754257,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 2.0,
18 "puct/buffer_construction_len/mean": 1751.625,
19 "puct/buffer_construction_len/std": 1643.474450695625,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7397.0,
22 "puct/sampled_value/mean": -1.510862661200017,
23 "puct/sampled_value/std": 0.0008662168905407264,
24 "puct/sampled_value/min": -1.5128135676775651,
25 "puct/sampled_value/max": -1.509671339880076,
26 "puct/sampled_timestep/mean": 2.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 2.0,
29 "puct/sampled_timestep/max": 2.0,
30 "puct/sampled_construction_len/mean": 1090.0,
31 "puct/sampled_construction_len/std": 155.88457268119896,
32 "puct/sampled_construction_len/min": 1000.0,
33 "puct/sampled_construction_len/max": 1360.0,
34 "time/sampling": 5213.487062454224,
35 "env/all/ac_tokens_per_turn": 7966.267578125,
36 "env/all/ob_tokens_per_turn": 2751.125,
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": 4078729,
41 "env/all/total_ob_tokens": 1408576,
42 "env/all/time/sampling_mean": 264.10583972744644,
43 "env/all/time/sampling_max": 350.07311153411865,
44 "env/all/time/env_step_mean": 2215.0262633892708,
45 "env/all/time/env_step_max": 4874.606095314026,
46 "env/all/reward/mean": 0.4587563424616886,
47 "env/all/reward/max": 0.6632904242765759,
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": 0.4587563424616886,
53 "env/all/correctness": 0.751953125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 2.218156834970131,
57 "env/all/raw_score/min": 1.5076352028717366,
58 "env/all/raw_score/max": 131.38251995130634,
59 "env/all/initial_raw_score": -1.5108626612000173,
60 "env/all/initial_raw_score/min": -1.5128135676775651,
61 "env/all/initial_raw_score/max": -1.509671339880076,
62 "env/all/msg": "Success; raw_score=1.9207633340749701",
63 "env/all/parsed_code": "```python\n\"\"\"Enhanced genetic and LP-guided search with adaptive direction and dynamic perturbations for C1 optimization.\"\"\"\n\nimport time\nimport numpy as np\nfrom scipy import optimize\nfrom scipy.optimize import minimize\nfrom scipy.optimize import linear_sum_assignment\n\nlinprog = optimize.linprog\n\ndef get_good_direction_to_move_into(sequence):\n \"\"\"Returns a better sequence using LP optimization, gradient descent, or adaptive perturbations.\"\"\"\n n = len(sequence)\n if n == 0:\n return None\n \n sum_sequence = np.sum(sequence)\n if sum_sequence <= 0.0:\n return None\n\n # Normalize sequence for LP\n normalized_sequence = [x * np.sqrt(2 * n) / sum_sequence for x in sequence]\n rhs = np.max(np.convolve(normalized_sequence, normalized_sequence))\n \n # Attempt to find better direction via LP\n g_fun = solve_convolution_lp(normalized_sequence, rhs)\n if g_fun is not None and np.sum(g_fun) > 0.0:\n # Scale back to original normalization\n normalized_g_fun = [x * np.sqrt(2 * n) / np.sum(g_fun) for x in g_fun]\n return normalized_g_fun\n\n # If LP fails, try gradient descent\n def objective(seq):\n return evaluate_sequence(seq)\n\n # Estimate gradient using finite differences\n eps = 1e-4\n grad = np.zeros(n)\n for i in range(n):\n perturbed_seq = [x + (eps if j == i else 0) for j, x in enumerate(sequence)]\n perturbed_seq = [max(0.0, x) for x in perturbed_seq]\n grad[i] = (objective(perturbed_seq) - objective(sequence)) / eps\n\n # Move in the direction of the negative gradient\n step_size = min(0.2, 0.1 / np.max(np.abs(grad)))\n new_seq = [x - step_size * grad[i] for i, x in enumerate(sequence)]\n new_seq = [max(0.0, x) for x in new_seq]\n return new_seq\n\ndef solve_convolution_lp(f_sequence, rhs):\n \"\"\"Solves LP to maximize sum(b) s.t. conv(f, b) <= rhs, b >= 0.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n # Construct constraint matrix and bounds\n c = -np.ones(n)\n a_ub = []\n b_ub = []\n\n for k in range(2 * n - 1):\n row = np.zeros(n)\n for i in range(n):\n j = k - i\n if 0 <= j < n:\n row[j] = f_sequence[i]\n a_ub.append(row)\n b_ub.append(rhs)\n\n # Non-negativity constraints\n a_ub_nonneg = -np.eye(n)\n b_ub_nonneg = np.zeros(n)\n a_ub = np.vstack([a_ub, a_ub_nonneg])\n b_ub = np.hstack([b_ub, b_ub_nonneg])\n\n # Solve LP with increased tolerances and time limit\n result = linprog(\n c,\n A_ub=a_ub,\n b_ub=b_ub,\n bounds=(0.0, None),\n method='highs',\n options={\"time_limit\": 30.0, \"feastol\": 1e-8, \"infodisp\": False}\n )\n if result.success:\n return result.x\n return None\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Use the last best construction if available\n prev = globals().get(\"GLOBAL_BEST_CONSTRUCTION\")\n if prev is not None and np.random.rand() < 0.5 and isinstance(prev, (list, tuple, np.ndarray)) and len(prev) > 0:\n best_sequence = list(np.asarray(prev, dtype=float).clip(0.0, 1000.0))\n else:\n # Use adaptive initialization: combine random and structured patterns\n length = 1000\n base_seq = np.sin(2 * np.pi * np.linspace(0, 1, length)) * 100.0 + 100.0\n noise = np.random.normal(0, 1, length)\n best_sequence = np.clip(base_seq + noise, 0.0, 1000.0).tolist()\n\n curr_sequence = best_sequence.copy()\n best_score = evaluate_sequence(curr_sequence)\n\n # Add additional diversity with sparse perturbations\n for i in range(5):\n # Add random sparse perturbation\n perturb = np.random.normal(0, 5.0, length)\n curr_sequence = [max(0.0, x + perturb[j]) for j, x in enumerate(curr_sequence)]\n try:\n curr_score = evaluate_sequence(curr_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = curr_sequence.copy()\n print(f\"New best score: {best_score}\")\n except Exception:\n print(\"Evaluation error, skipping update\")\n\n while time.time() < deadline:\n # Attempt to find a new direction\n h_sequence = get_good_direction_to_move_into(curr_sequence)\n if not h_sequence:\n print(\" Unable to find new direction, applying adaptive structured perturbation.\")\n # Apply structured perturbation: combine sine and random waves\n perturb = np.sin(2 * np.pi * np.random.rand(length) / length) * 10.0\n for i in range(len(curr_sequence)):\n curr_sequence[i] = max(0.0, curr_sequence[i] + perturb[i])\n else:\n curr_sequence = h_sequence\n\n try:\n curr_score = evaluate_sequence(curr_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = curr_sequence.copy()\n print(f\"New best score: {best_score}\")\n except Exception:\n print(\"Evaluation error, skipping update\")\n\n return [float(max(0.0, x)) for x in best_sequence]\n```",
64 "env/all/time/policy": 264.10583972744644,
65 "env/all/time/policy/min": 69.49347066879272,
66 "env/all/time/policy/max": 350.07311153411865,
67 "env/all/time/env_step": 2215.0262633892708,
68 "env/all/time/env_step/min": 0.0074422359466552734,
69 "env/all/time/env_step/max": 4874.606095314026,
70 "env/all/time/reward_compute": 9.564682841300964e-07,
71 "env/all/time/reward_compute/min": 1.862645149230957e-07,
72 "env/all/time/reward_compute/max": 3.2186508178710938e-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": 0.027251766994595528,
77 "advantage/min": -1.0,
78 "advantage/max": 20.6398983001709,
79 "time/assemble_training_data": 9.06363844871521,
80 "time/kl_vs_base": 82.385498046875,
81 "kl_policy_base": 0.0006102915504015982,
82 "time/train": 502.71834444999695,
83 "time/save_checkpoint": 33.26317811012268,
84 "time/total": 5842.74027466774
85}[2026-07-09T06:24:18+00:00] job=1812628 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:30:32+00:00] job=1813127 node=node-22 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:07:08+00:00] job=1813128 node=node-6 ngpu=6 ntrain=2 replicas=4 flash_attn=yes