Views
No views yet
ac1. Checkpoint saved
after training step 7 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 7,
3 "progress/batch": 7,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.16,
6 "puct/buffer_size": 120,
7 "puct/sampled_size": 8,
8 "puct/T": 3584,
9 "puct/scale_last": 0.49195678802302667,
10 "puct/buffer_value/mean": -1.5768942279196685,
11 "puct/buffer_value/std": 0.15182478014639553,
12 "puct/buffer_value/min": -2.000000000000007,
13 "puct/buffer_value/max": -1.5063036919514743,
14 "puct/buffer_timestep/mean": 2.7333333333333334,
15 "puct/buffer_timestep/std": 2.174600857373345,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 6.0,
18 "puct/buffer_construction_len/mean": 1499.9166666666667,
19 "puct/buffer_construction_len/std": 1438.6672801782752,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7850.0,
22 "puct/sampled_value/mean": -1.507116542131942,
23 "puct/sampled_value/std": 0.00046953853252642354,
24 "puct/sampled_value/min": -1.5074196007171705,
25 "puct/sampled_value/max": -1.5063036919514743,
26 "puct/sampled_timestep/mean": 6.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 6.0,
29 "puct/sampled_timestep/max": 6.0,
30 "puct/sampled_construction_len/mean": 1000.0,
31 "puct/sampled_construction_len/std": 0.0,
32 "puct/sampled_construction_len/min": 1000.0,
33 "puct/sampled_construction_len/max": 1000.0,
34 "time/sampling": 5193.891042709351,
35 "env/all/ac_tokens_per_turn": 7896.609375,
36 "env/all/ob_tokens_per_turn": 3011.0,
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": 4043064,
41 "env/all/total_ob_tokens": 1541632,
42 "env/all/time/sampling_mean": 268.3592103184201,
43 "env/all/time/sampling_max": 347.3218319416046,
44 "env/all/time/env_step_mean": 2071.404621460475,
45 "env/all/time/env_step_max": 4852.551887512207,
46 "env/all/reward/mean": 0.49589447348375393,
47 "env/all/reward/max": 0.663876745907893,
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": 0.49589447348375393,
53 "env/all/correctness": 0.759765625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 1.7077152923523793,
57 "env/all/raw_score/min": 1.5063036919506345,
58 "env/all/raw_score/max": 69.29544253326463,
59 "env/all/initial_raw_score": -1.507116542131942,
60 "env/all/initial_raw_score/min": -1.5074196007171705,
61 "env/all/initial_raw_score/max": -1.5063036919514743,
62 "env/all/msg": "Success; raw_score=1.550797803997189",
63 "env/all/parsed_code": "```python\n# EVOLVE-BLOCK-START\n\"\"\"Enhanced Multi-Strategy Optimization with Adaptive Perturbation and Constraint Expansion for C1 Autoconvolution Minimization\"\"\"\nimport time\nimport numpy as np\nfrom scipy.optimize import linprog\nimport copy\nimport random\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Start from known best construction if it exists\n prev = globals().get(\"height_sequence_1\", None)\n if prev is not None and isinstance(prev, list) and len(prev) > 0:\n best_sequence = list(np.asarray(prev, dtype=float))\n else:\n n = 1000\n # Generate a peak-focused initial sequence\n center = n // 2\n base = np.zeros(n)\n for i in range(n):\n distance = abs(i - center)\n base[i] = 0.3 * (1 - distance / center) if distance < center else 0.0\n # Add some randomness\n gauss = np.random.normal(0, 0.05, n)\n best_sequence = [max(0.0, base[i] + gauss[i]) for i in range(n)]\n \n current_sequence = best_sequence.copy()\n best_score = evaluate_sequence(current_sequence)\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using advanced LP with adaptive constraints and optimized step size.\"\"\"\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 normalized_sequence = [x * np.sqrt(2 * n) / sum_sequence for x in sequence]\n conv = np.convolve(normalized_sequence, normalized_sequence)\n max_b = np.max(conv)\n\n # Generate tight positions by selecting top K positions\n K = min(100, len(conv)) # select top 100 positions\n tight_positions = np.argsort(conv)[-K:]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n\n # Solve LP with extended time limit\n g_fun = solve_convolution_lp(normalized_sequence, max_b, tight_positions)\n if g_fun is None:\n return None\n\n sum_g = np.sum(g_fun)\n if sum_g <= 0.0:\n return None\n\n normalized_g = [x * np.sqrt(2 * n) / sum_g for x in g_fun]\n\n # Adaptive line search with refined t values using step decay\n best_t = 0.01\n best_score = float('inf')\n # Use a greedy step decay approach\n for t in [0.9, 0.5, 0.2, 0.1, 0.05, 0.02, 0.01]:\n new_sequence = [(1 - t) * x + t * y for x, y in zip(sequence, normalized_g)]\n try:\n curr_score = evaluate_sequence(new_sequence)\n if curr_score < best_score:\n best_t = t\n best_score = curr_score\n except Exception:\n continue\n\n return [(1 - best_t) * x + best_t * y for x, y in zip(sequence, normalized_g)]\n\n def solve_convolution_lp(f_sequence, rhs, tight_positions):\n \"\"\"Solves LP with tight positions and extended time limit.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n c = -np.ones(n)\n a_ub = []\n b_ub = []\n for k in tight_positions:\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 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 result = linprog(\n c,\n A_ub=a_ub,\n b_ub=b_ub,\n bounds=[(0, 1000.0) for _ in range(n)],\n options={\n \"time_limit\": 60.0, # Increased time limit\n \"disp\": False,\n },\n )\n if result.success:\n return result.x\n return None\n\n def perturb_sequence(seq):\n \"\"\"Perturbation with focus on contributing indices to max convolution.\"\"\"\n conv = np.convolve(seq, seq)\n max_conv_pos = np.argmax(conv)\n contributing_indices = set()\n for i in range(len(seq)):\n j = max_conv_pos - i\n if 0 <= j < len(seq):\n contributing_indices.add(i)\n contributing_indices.add(j)\n if not contributing_indices:\n # Fallback to random perturbation\n indices = np.random.choice(len(seq), size=2, replace=False)\n else:\n indices = np.random.choice(list(contributing_indices), size=2, replace=False)\n new_seq = seq.copy()\n for idx in indices:\n new_val = max(0.0, new_seq[idx] + np.random.normal(0, 0.05))\n new_seq[idx] = new_val\n return new_seq\n\n while time.time() < deadline:\n h_function = get_good_direction_to_move_into(current_sequence)\n if h_function is None:\n current_sequence = perturb_sequence(current_sequence)\n else:\n current_sequence = h_function\n\n try:\n curr_score = evaluate_sequence(current_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = current_sequence.copy()\n print(f\"New best: {best_score}\")\n except Exception:\n pass\n\n return [float(max(0.0, x)) for x in best_sequence]\n# EVOLVE-BLOCK-END\n```",
64 "env/all/time/policy": 268.3592103184201,
65 "env/all/time/policy/min": 125.87504696846008,
66 "env/all/time/policy/max": 347.3218319416046,
67 "env/all/time/env_step": 2071.404621460475,
68 "env/all/time/env_step/min": 0.005207538604736328,
69 "env/all/time/env_step/max": 4852.551887512207,
70 "env/all/time/reward_compute": 3.310851752758026e-07,
71 "env/all/time/reward_compute/min": 2.2724270820617676e-07,
72 "env/all/time/reward_compute/max": 8.009374141693115e-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": 0.027937524020671844,
77 "advantage/min": -1.0,
78 "advantage/max": 17.481287002563477,
79 "time/assemble_training_data": 8.66896677017212,
80 "time/kl_vs_base": 87.524334192276,
81 "kl_policy_base": 0.0007810262613929808,
82 "time/train": 516.926687002182,
83 "time/save_checkpoint": 12.586852312088013,
84 "time/total": 5821.849282503128
85}[2026-07-09T06:24:18+00:00] job=1812630 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T09:14:17+00:00] job=1813129 node=node-28 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:46:36+00:00] job=1813130 node=node-22 ngpu=6 ntrain=2 replicas=4 flash_attn=yes