Views
No views yet
ac1. Checkpoint saved
after training step 15 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 15,
3 "progress/batch": 15,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.32,
6 "puct/buffer_size": 248,
7 "puct/sampled_size": 8,
8 "puct/T": 7680,
9 "puct/scale_last": 0.3992471016406822,
10 "puct/buffer_value/mean": -1.5326054369725994,
11 "puct/buffer_value/std": 0.09806630942901592,
12 "puct/buffer_value/min": -2.000000000000008,
13 "puct/buffer_value/max": -1.5063759391533882,
14 "puct/buffer_timestep/mean": 6.741935483870968,
15 "puct/buffer_timestep/std": 4.479110956828331,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 14.0,
18 "puct/buffer_construction_len/mean": 1175.5282258064517,
19 "puct/buffer_construction_len/std": 841.8743292796604,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7397.0,
22 "puct/sampled_value/mean": -1.5063859363810783,
23 "puct/sampled_value/std": 8.07889009865774e-06,
24 "puct/sampled_value/min": -1.5064045942034718,
25 "puct/sampled_value/max": -1.5063759391533882,
26 "puct/sampled_timestep/mean": 14.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 14.0,
29 "puct/sampled_timestep/max": 14.0,
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": 6211.373323440552,
35 "env/all/ac_tokens_per_turn": 8530.6796875,
36 "env/all/ob_tokens_per_turn": 2997.0,
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": 4367708,
41 "env/all/total_ob_tokens": 1534464,
42 "env/all/time/sampling_mean": 297.56582681508735,
43 "env/all/time/sampling_max": 385.2249779701233,
44 "env/all/time/env_step_mean": 3004.890956228599,
45 "env/all/time/env_step_max": 5833.007267713547,
46 "env/all/reward/mean": 0.6155119335249499,
47 "env/all/reward/max": 0.663845150642878,
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": 0.6155119335249499,
53 "env/all/correctness": 0.953125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 6.132365471197116,
57 "env/all/raw_score/min": 1.506375383465757,
58 "env/all/raw_score/max": 2000.0,
59 "env/all/initial_raw_score": -1.5063859363810783,
60 "env/all/initial_raw_score/min": -1.5064045942034718,
61 "env/all/initial_raw_score/max": -1.5063759391533882,
62 "env/all/msg": "RuntimeError: Program execution failed: TypeError: only integer scalar arrays can be converted to a scalar index",
63 "env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy import optimize\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score \n using enhanced LP-based directions, adaptive blending with line search, and\n diversified initialization to avoid local minima. Focuses on efficiency and\n convergence by reducing LP constraints and refining perturbation strategies.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Initialize using known sequences or diverse patterns\n n = 1000\n best_sequence = None\n\n # Diverse initializations: uniform, Gaussian, and sparse patterns\n def create_initial_sequence():\n seq1 = np.random.normal(0.5, 0.15, n)\n seq1 = np.max(0.0, np.clip(seq1, 0, 1000.0)) * np.sqrt(2 * n) / np.sum(seq1)\n seq2 = np.ones(n) * np.sqrt(2 * n) / n\n seq3 = np.zeros(n)\n seq3[0] = np.sqrt(2 * n)\n seq3 = np.max(0.0, np.clip(seq3, 0, 1000.0))\n return [0.7 * seq1[i] + 0.2 * seq2[i] + 0.1 * seq3[i] for i in range(n)]\n\n # Try all initializations\n for _ in range(3):\n candidate = create_initial_sequence()\n score = evaluate_sequence(candidate)\n if best_sequence is None or score < evaluate_sequence(best_sequence):\n best_sequence = candidate\n\n curr_sequence = [float(x) for x in best_sequence]\n best_score = evaluate_sequence(curr_sequence)\n\n # Parameters for optimization\n perturb_rate = 0.05\n blend_factor = 0.8\n min_perturb = 0.005\n\n # Helper to solve LP with single constraint\n def solve_convolution_lp(f_sequence, rhs, max_idx):\n n = len(f_sequence)\n if n == 0:\n return None\n c = -np.ones(n)\n a_ub = []\n b_ub = []\n row = np.zeros(n)\n for i in range(n):\n j = max_idx - i\n if 0 <= j < n:\n row[j] = f_sequence[i]\n a_ub.append(row)\n b_ub.append(rhs)\n try:\n result = optimize.linprog(\n c,\n A_ub=a_ub,\n b_ub=b_ub,\n bounds=(0.0, 1000.0),\n method='highs',\n options={\n \"time_limit\": 300.0,\n \"disp\": False,\n },\n )\n if result.success:\n return result.x\n except:\n pass\n return None\n\n while time.time() < deadline:\n try:\n # Compute current convolution and max_idx\n conv = np.convolve(curr_sequence, curr_sequence)\n max_idx = np.argmax(conv)\n sum_a = np.sum(curr_sequence)\n if sum_a < 0.01:\n raise ValueError(\"Sum too small\")\n\n # Normalize for LP\n normalized_seq = [x / sum_a * np.sqrt(2 * n) for x in curr_sequence]\n rhs = conv[max_idx]\n\n # Solve LP with single constraint\n g_fun = solve_convolution_lp(normalized_seq, rhs, max_idx)\n if g_fun is not None and np.sum(g_fun) > 0.0:\n sum_g = np.sum(g_fun)\n normalized_g = [x / sum_g * np.sqrt(2 * n) for x in g_fun]\n\n # Line search for optimal blend\n def objective(t):\n new_seq = [(1 - t) * x + t * y for x, y in zip(curr_sequence, normalized_g)]\n return evaluate_sequence(new_seq)\n\n result = optimize.minimize_scalar(objective, bounds=(0.0, 1.0), method='bounded', tol=1e-5)\n best_t = result.x\n new_seq = [(1 - best_t) * x + best_t * y for x, y in zip(curr_sequence, normalized_g)]\n curr_sequence = new_seq\n else:\n # Fallback to perturbation on top contributors\n contributions = []\n for i in range(n):\n j = max_idx - i\n if 0 <= j < n:\n contributions.append((i, abs(curr_sequence[i] * curr_sequence[j])))\n if contributions:\n sorted_contributions = sorted(contributions, key=lambda x: x[1], reverse=True)\n idxs = [sorted_contributions[i][0] for i in range(min(10, len(sorted_contributions)))]\n curr_sequence = perturb_sequence(curr_sequence, idxs, perturb_rate)\n else:\n idx = np.random.randint(0, n)\n curr_sequence = perturb_sequence(curr_sequence, [idx], perturb_rate)\n\n except Exception as e:\n print(f\"Error during optimization: {e}. Skipping update.\")\n continue\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\ndef perturb_sequence(sequence, idxs, perturb_amount):\n \"\"\"Perturb specified indices in sequence with controlled amounts.\"\"\"\n new_seq = sequence.copy()\n for idx in idxs:\n new_seq[idx] = max(0.0, new_seq[idx] - perturb_amount * new_seq[idx])\n new_seq[idx] = min(1000.0, new_seq[idx])\n return new_seq\n```",
64 "env/all/time/policy": 297.56582681508735,
65 "env/all/time/policy/min": 105.472647190094,
66 "env/all/time/policy/max": 385.2249779701233,
67 "env/all/time/env_step": 3004.890956228599,
68 "env/all/time/env_step/min": 0.05785202980041504,
69 "env/all/time/env_step/max": 5833.007267713547,
70 "env/all/time/reward_compute": 3.3760443329811096e-07,
71 "env/all/time/reward_compute/min": 2.123415470123291e-07,
72 "env/all/time/reward_compute/max": 8.568167686462402e-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": 0.029155215248465538,
77 "advantage/min": -1.0,
78 "advantage/max": 16.547401428222656,
79 "time/assemble_training_data": 9.759180784225464,
80 "time/kl_vs_base": 92.90454077720642,
81 "kl_policy_base": 0.0007622624398209155,
82 "time/train": 549.8254308700562,
83 "time/save_checkpoint": 34.76782250404358,
84 "time/total": 6900.638512372971
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