Views
No views yet
ac1. Checkpoint saved
after training step 30 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 30,
3 "progress/batch": 30,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.62,
6 "puct/buffer_size": 488,
7 "puct/sampled_size": 8,
8 "puct/T": 15360,
9 "puct/scale_last": 0.4936516100901467,
10 "puct/buffer_value/mean": -1.521551222181111,
11 "puct/buffer_value/std": 0.07532092758477031,
12 "puct/buffer_value/min": -2.000000000000008,
13 "puct/buffer_value/max": -1.5063483899098538,
14 "puct/buffer_timestep/mean": 14.245901639344263,
15 "puct/buffer_timestep/std": 8.806956523308148,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 29.0,
18 "puct/buffer_construction_len/mean": 1089.202868852459,
19 "puct/buffer_construction_len/std": 606.5361559168529,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7397.0,
22 "puct/sampled_value/mean": -1.5063602553675353,
23 "puct/sampled_value/std": 3.984250278562637e-12,
24 "puct/sampled_value/min": -1.506360255371331,
25 "puct/sampled_value/max": -1.5063602553592976,
26 "puct/sampled_timestep/mean": 29.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 29.0,
29 "puct/sampled_timestep/max": 29.0,
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": 6103.4288284778595,
35 "env/all/ac_tokens_per_turn": 8228.41015625,
36 "env/all/ob_tokens_per_turn": 3024.25,
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": 4212946,
41 "env/all/total_ob_tokens": 1548416,
42 "env/all/time/sampling_mean": 283.88724840944633,
43 "env/all/time/sampling_max": 377.5336437225342,
44 "env/all/time/env_step_mean": 2940.9875311735086,
45 "env/all/time/env_step_max": 5720.159309864044,
46 "env/all/reward/mean": 0.6386215891149316,
47 "env/all/reward/max": 0.6638519065017691,
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": 0.6386215891149316,
53 "env/all/correctness": 0.98046875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 1.5422631534514308,
57 "env/all/raw_score/min": 1.5063600534508914,
58 "env/all/raw_score/max": 2.0893790637193503,
59 "env/all/initial_raw_score": -1.506360255367535,
60 "env/all/initial_raw_score/min": -1.506360255371331,
61 "env/all/initial_raw_score/max": -1.5063602553592976,
62 "env/all/msg": "Success; raw_score=1.5063602553592976",
63 "env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy.optimize import linprog\nimport random\nimport copy\nimport math\n\ndef propose_candidate(seed=42, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score \n using a combination of symmetry-based initialization, adaptive constraint selection,\n and enhanced line search strategies. This approach aims to find a structured initial \n configuration to avoid local optima while exploring the solution space with\n more sophisticated scaling and combination strategies.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + 1000 - 10 # Run for up to 1000 seconds\n\n n = 1000\n # Start with a symmetric Gaussian-like initialization\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 # Initial sequence: Gaussian-like symmetric\n center = n // 2\n sigma = 100\n best_sequence = np.zeros(n)\n for i in range(n):\n dist = abs(i - center)\n val = np.exp(-dist**2 / (2 * sigma**2))\n best_sequence[i] = val\n # Normalize to ensure sum is at least 0.01\n sum_a = np.sum(best_sequence)\n if sum_a < 0.001:\n best_sequence = np.full(n, 1.0)\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 # Parameters for line search\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, mode='full')\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 # Dynamically select active constraints\n K = n # Use all indices\n conv_sorted_indices = np.argsort(conv)\n active_indices = conv_sorted_indices[-K:]\n\n # Build the A matrix for LP constraint system\n A = []\n for j in active_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(active_indices), max_b)\n\n # Define LP problem: maximize sum of g_0 subject to constraints\n c = [-1.0] * n\n bounds = [(0.0, 10000.0) for _ in range(n)] # Use higher upper bound\n\n # Solve the LP using the 'highs' method\n res = linprog(c=c, A_ub=A, b_ub=b, bounds=bounds, method='highs', options={\"feastol\": 1e-6})\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 # Perform line search for scale with wider range\n def score_func(scale):\n scaled_g = g_0 * scale\n new_sequence = scaled_g\n return evaluate_sequence(new_sequence.tolist())\n\n a, b = 0.1, 10.0 # Wider range for scale\n iterations = 150\n for _ in range(iterations):\n c_val = a + (b - a)/2\n d_val = a + (b - a)*0.381966\n f_c = score_func(c_val)\n f_d = score_func(d_val)\n if f_d < f_c:\n b = c_val\n else:\n a = d_val\n best_scale = (a + b)/2\n g_0_normalized = g_0 * best_scale\n g_0_normalized = np.clip(g_0_normalized, 0.0, 1000.0)\n\n # Define objective function for line search to find optimal alpha\n def score_func_alpha(alpha):\n new_sequence = (1 - alpha) * curr_sequence + alpha * g_0_normalized\n new_sequence = np.clip(new_sequence, 0.0, 1000.0)\n return evaluate_sequence(new_sequence.tolist())\n\n # Use golden-section search to find the optimal alpha between 0 and 1\n a, b = 0.0, 1.0\n iterations = 150\n for _ in range(iterations):\n c_val = a + (b - a)/2\n d_val = a + (b - a)*0.381966\n f_c = score_func_alpha(c_val)\n f_d = score_func_alpha(d_val)\n if f_d < f_c:\n b = c_val\n else:\n a = d_val\n best_alpha = (a + b)/2\n\n # Apply the best alpha found\n curr_sequence = (1 - best_alpha) * curr_sequence + best_alpha * g_0_normalized\n curr_sequence = np.clip(curr_sequence, 0.0, 1000.0)\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(\"LP solution not found, breaking\")\n break\n\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 > 100: # Reduced threshold for earlier stopping\n print(\"No improvement for 100 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 to ensure non-negative and valid values\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": 283.88724840944633,
65 "env/all/time/policy/min": 111.25921845436096,
66 "env/all/time/policy/max": 377.5336437225342,
67 "env/all/time/env_step": 2940.9875311735086,
68 "env/all/time/env_step/min": 2.5509610176086426,
69 "env/all/time/env_step/max": 5720.159309864044,
70 "env/all/time/reward_compute": 3.7299469113349915e-07,
71 "env/all/time/reward_compute/min": 2.5331974029541016e-07,
72 "env/all/time/reward_compute/max": 8.083879947662354e-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": 0.004564481787383556,
77 "advantage/min": -1.0,
78 "advantage/max": 0.615384578704834,
79 "time/assemble_training_data": 9.993045330047607,
80 "time/kl_vs_base": 94.99324822425842,
81 "kl_policy_base": 0.0007551112212240696,
82 "time/train": 532.2399046421051,
83 "time/save_checkpoint": 34.483798027038574,
84 "time/total": 6779.2794778347015
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