Views
No views yet
ac1. Checkpoint saved
after training step 45 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 45,
3 "progress/batch": 45,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.92,
6 "puct/buffer_size": 728,
7 "puct/sampled_size": 8,
8 "puct/T": 23040,
9 "puct/scale_last": 251.17629545984107,
10 "puct/buffer_value/mean": -1.873608266638335,
11 "puct/buffer_value/std": 9.302536570864394,
12 "puct/buffer_value/min": -252.6826286848707,
13 "puct/buffer_value/max": -1.5063332250296213,
14 "puct/buffer_timestep/mean": 21.747252747252748,
15 "puct/buffer_timestep/std": 13.136304368147167,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 44.0,
18 "puct/buffer_construction_len/mean": 1059.7953296703297,
19 "puct/buffer_construction_len/std": 498.3607200471164,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7397.0,
22 "puct/sampled_value/mean": -1.506333225029624,
23 "puct/sampled_value/std": 1.6050308076832566e-15,
24 "puct/sampled_value/min": -1.506333225029626,
25 "puct/sampled_value/max": -1.5063332250296215,
26 "puct/sampled_timestep/mean": 43.625,
27 "puct/sampled_timestep/std": 0.9921567416492215,
28 "puct/sampled_timestep/min": 41.0,
29 "puct/sampled_timestep/max": 44.0,
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": 3286.2794048786163,
35 "env/all/ac_tokens_per_turn": 8716.423828125,
36 "env/all/ob_tokens_per_turn": 2715.5,
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": 4462809,
41 "env/all/total_ob_tokens": 1390336,
42 "env/all/time/sampling_mean": 302.14945559902117,
43 "env/all/time/sampling_max": 403.9773061275482,
44 "env/all/time/env_step_mean": 1241.9229878797196,
45 "env/all/time/env_step_max": 2887.0948646068573,
46 "env/all/reward/mean": 0.6278732243731155,
47 "env/all/reward/max": 0.6638637299802627,
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": 0.6278732243731155,
53 "env/all/correctness": 0.955078125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 1.5259581284406505,
57 "env/all/raw_score/min": 1.5063332250296213,
58 "env/all/raw_score/max": 2.371067926216342,
59 "env/all/initial_raw_score": -1.506333225029624,
60 "env/all/initial_raw_score/min": -1.506333225029626,
61 "env/all/initial_raw_score/max": -1.5063332250296215,
62 "env/all/msg": "Success; raw_score=1.5063332250296215",
63 "env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy.optimize import linprog, minimize\n\ndef propose_candidate(seed=42, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score \n through a combination of multi-objective optimization with enhanced exploration.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + 1000\n\n # Use the existing best sequence if available, default to a single-spike initial guess\n if 'height_sequence_1' in globals() and isinstance(height_sequence_1, (list, np.ndarray)):\n best_sequence = np.array(list(height_sequence_1))\n else:\n # Start with a single-spike sequence\n n = 1000\n initial_sequence = np.zeros(n)\n initial_sequence[500] = np.sqrt(2 * n)\n best_sequence = initial_sequence.copy()\n \n curr_sequence = best_sequence.copy()\n best_score = evaluate_sequence(curr_sequence.tolist())\n no_improvement_counter = 0\n iteration_count = 0\n n = len(curr_sequence)\n\n # Parameters for the optimization steps\n k_values = [500, 700, 900] # Larger number of constraints for better tightness\n perturbation_scale = 0.1 # Increased perturbation scale for exploration\n alpha_search_iterations = 100\n\n # Track the maximum index for tighter constraints\n while time.time() < deadline:\n current_time = time.time()\n try:\n # Compute current convolution and max_b\n conv = np.convolve(curr_sequence, curr_sequence)\n max_b = np.max(conv)\n sum_a = np.sum(curr_sequence)\n if sum_a < 0.001:\n raise ValueError(\"Sum too small\")\n\n # Select dynamic constraints based on current convolution\n k = k_values[iteration_count % len(k_values)]\n sorted_conv = sorted(enumerate(conv), key=lambda x: -x[1])\n top_indices = [i for i, _ in sorted_conv[:k]]\n A = []\n for j in top_indices:\n row = np.zeros(n)\n for m in range(n):\n i = j - m\n if 0 <= i < n:\n row[m] = curr_sequence[i]\n A.append(row)\n A = np.array(A)\n b = np.full(len(top_indices), max_b * 0.8) # Aggressive constraint tightening\n\n # Define LP problem: maximize sum of g_0 subject to constraints\n c = [-1.0] * n\n bounds = [(0.0, 1000.0) for _ in range(n)]\n\n # Solve the LP using 'highs' for performance\n res = linprog(c=c, A_ub=A, b_ub=b, bounds=bounds, method='highs', options={\"feastol\": 1e-8})\n if res.success:\n g_0 = res.x\n sum_g = np.sum(g_0)\n if sum_g == 0:\n continue # Avoid division by zero\n\n # Normalize g_0 to match the normalization\n g_0_normalized = g_0 / sum_g * np.sqrt(2 * n)\n\n # Define function to minimize for alpha search, using a better optimization\n def evaluate_alpha(alpha):\n new_sequence = (1 - alpha) * curr_sequence + alpha * g_0_normalized\n conv_new = np.convolve(new_sequence, new_sequence)\n new_max_b = np.max(conv_new)\n sum_new = np.sum(new_sequence)\n return 2 * n * new_max_b / (sum_new ** 2) # Simplified for line search\n\n # Use scipy's minimize with L-BFGS-B for more accurate alpha search\n result = minimize(evaluate_alpha, x0=0.5, method='L-BFGS-B', bounds=[(0.0, 1.0)])\n if result.success:\n alpha_opt = result.x[0]\n new_sequence = (1 - alpha_opt) * curr_sequence + alpha_opt * g_0_normalized\n\n # Ensure non-negative and bounded values\n new_sequence = np.clip(new_sequence, 0.0, 1000.0)\n new_sequence = new_sequence / np.sum(new_sequence) * np.sqrt(2 * n)\n curr_sequence = new_sequence.copy()\n\n # Evaluate and update best sequence\n curr_score = evaluate_sequence(curr_sequence.tolist())\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = curr_sequence.copy()\n no_improvement_counter = 0\n print(f\"New best score: {best_score}\")\n else:\n no_improvement_counter += 1\n else:\n print(\"Alpha optimization not found, using random perturbation\")\n # Random perturbation if alpha optimization fails\n perturb = np.random.normal(0, perturbation_scale, size=n)\n curr_sequence = np.clip(curr_sequence + perturb, 0.0, 1000.0)\n curr_sequence = curr_sequence / np.sum(curr_sequence) * np.sqrt(2 * n)\n\n else:\n print(\"LP solution not found, trying random perturbation\")\n # Random perturbation if LP fails\n perturb = np.random.normal(0, perturbation_scale, size=n)\n curr_sequence = np.clip(curr_sequence + perturb, 0.0, 1000.0)\n curr_sequence = curr_sequence / np.sum(curr_sequence) * np.sqrt(2 * n)\n\n # Early stopping if no improvement for many iterations\n iteration_count += 1\n if iteration_count % 100 == 0:\n print(f\"Iteration {iteration_count}, current best score: {best_score}\")\n if no_improvement_counter > 200:\n print(\"No improvement for 200 iterations, stopping early\")\n break\n\n except Exception as e:\n print(f\"Error during optimization: {e}. Skipping update.\")\n continue\n\n # Final safety check and return\n final_sequence = np.maximum(0.0, best_sequence).tolist()\n return [float(x) for x in final_sequence]\n```",
64 "env/all/time/policy": 302.14945559902117,
65 "env/all/time/policy/min": 115.7823326587677,
66 "env/all/time/policy/max": 403.9773061275482,
67 "env/all/time/env_step": 1241.9229878797196,
68 "env/all/time/env_step/min": 1.4934437274932861,
69 "env/all/time/env_step/max": 2887.0948646068573,
70 "env/all/time/reward_compute": 7.287599146366119e-07,
71 "env/all/time/reward_compute/min": 2.5704503059387207e-07,
72 "env/all/time/reward_compute/max": 2.7567148208618164e-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": 0.0016624608542770147,
77 "advantage/min": -1.0,
78 "advantage/max": 0.36956512928009033,
79 "time/assemble_training_data": 9.778876066207886,
80 "time/kl_vs_base": 97.5128424167633,
81 "kl_policy_base": 0.0007102805539034307,
82 "time/train": 547.6965055465698,
83 "time/save_checkpoint": 10.63455605506897,
84 "time/total": 3955.327879667282
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
[2026-07-11T13:13:37+00:00] job=1824165 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-12T04:58:24+00:00] job=1827204 node=node-4 ngpu=6 ntrain=2 replicas=4 flash_attn=yes