Views
No views yet
ac1. Checkpoint saved
after training step 32 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 32,
3 "progress/batch": 32,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.66,
6 "puct/buffer_size": 520,
7 "puct/sampled_size": 8,
8 "puct/T": 16384,
9 "puct/scale_last": 0.4936516100901467,
10 "puct/buffer_value/mean": -1.5206163767340508,
11 "puct/buffer_value/std": 0.07305783369158816,
12 "puct/buffer_value/min": -2.000000000000008,
13 "puct/buffer_value/max": -1.5063483899098538,
14 "puct/buffer_timestep/mean": 15.246153846153845,
15 "puct/buffer_timestep/std": 9.384161401377355,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 31.0,
18 "puct/buffer_construction_len/mean": 1083.7134615384616,
19 "puct/buffer_construction_len/std": 587.9681118802428,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7397.0,
22 "puct/sampled_value/mean": -1.5063591938235639,
23 "puct/sampled_value/std": 2.0790019123920306e-06,
24 "puct/sampled_value/min": -1.5063602553528839,
25 "puct/sampled_value/max": -1.5063540432739218,
26 "puct/sampled_timestep/mean": 31.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 31.0,
29 "puct/sampled_timestep/max": 31.0,
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": 5512.087729692459,
35 "env/all/ac_tokens_per_turn": 8117.90625,
36 "env/all/ob_tokens_per_turn": 3077.625,
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": 4156368,
41 "env/all/total_ob_tokens": 1575744,
42 "env/all/time/sampling_mean": 489.5367934964597,
43 "env/all/time/sampling_max": 683.1813368797302,
44 "env/all/time/env_step_mean": 2623.7749838302843,
45 "env/all/time/env_step_max": 4826.177067518234,
46 "env/all/reward/mean": 0.6058640082281155,
47 "env/all/reward/max": 0.663861419857056,
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": 0.6058640082281155,
53 "env/all/correctness": 0.9296875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 1.5453690119007386,
57 "env/all/raw_score/min": 1.5063384668093949,
58 "env/all/raw_score/max": 3.2929185555542486,
59 "env/all/initial_raw_score": -1.506359193823564,
60 "env/all/initial_raw_score/min": -1.5063602553528839,
61 "env/all/initial_raw_score/max": -1.5063540432739218,
62 "env/all/msg": "Success; raw_score=1.506340751346248",
63 "env/all/parsed_code": "```python\nimport numpy as np\nimport time\nfrom scipy.optimize import linprog\nimport random\nimport copy\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 structured initialization, adaptive constraint selection, and enhanced exploration.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + 1000 - 10 # Run for up to 1000 seconds\n\n # Use the existing best sequence if available\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 # Initialize with a uniform distribution scaled to the required sum\n n = 1000\n best_sequence = np.ones(n)\n best_sequence = best_sequence * np.sqrt(2 * n) / np.sum(best_sequence)\n if n % 2 == 1:\n best_sequence[n//2] *= 2 # Slight asymmetry to avoid perfect symmetry\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 initial_sum = np.sum(curr_sequence)\n\n # Parameters for adaptive constraint selection and line search\n k_values = [500, 700, 900, 1000] # Increased number of constraints for diversity\n perturbation_freq = 75 # Increase frequency of random perturbations\n max_iterations = 2000 # Increased iteration limit for more exploration\n\n while time.time() < deadline and iteration_count < max_iterations:\n current_time = time.time()\n iteration_count += 1\n\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 constraints using a mix of top indices and random indices\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 random_indices = np.random.choice(range(len(conv)), size=50, replace=False)\n top_indices = np.unique(np.concatenate([top_indices, random_indices])).tolist()\n A = []\n for j in top_indices:\n row = np.zeros(1000)\n for m in range(1000):\n i = j - m\n if 0 <= i < 1000:\n row[m] = curr_sequence[i]\n A.append(row)\n A = np.array(A)\n b = np.full(len(top_indices), max_b - 0.005) # Slightly lower bound to challenge optimization\n\n # Define LP problem: maximize sum of g_0 subject to constraints\n c = [-1.0] * 1000\n bounds = [(0.0, 1000.0) for _ in range(1000)]\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 g_0_normalized = g_0 / sum_g * sum_a\n\n # Perform adaptive line search with fewer iterations for speed\n def find_optimal_alpha(curr_seq, g0_norm):\n low = 0.0\n high = 1.0\n for _ in range(500): # Reduced iterations for speed\n m1 = low + (high - low)/3\n m2 = high - (high - low)/3\n new_seq1 = (1 - m1) * curr_seq + m1 * g0_norm\n new_seq2 = (1 - m2) * curr_seq + m2 * g0_norm\n conv_new1 = np.convolve(new_seq1, new_seq1)\n conv_new2 = np.convolve(new_seq2, new_seq2)\n new_max_b1 = np.max(conv_new1)\n new_max_b2 = np.max(conv_new2)\n \n if new_max_b1 < new_max_b2:\n high = m2\n else:\n low = m1\n alpha_opt = (low + high) / 2\n new_seq = (1 - alpha_opt) * curr_seq + alpha_opt * g0_norm\n return new_seq, np.max(np.convolve(new_seq, new_seq))\n \n new_sequence, new_score = find_optimal_alpha(curr_sequence, g_0_normalized)\n \n # Relax the max_b check to allow more updates\n conv_new = np.convolve(new_sequence, new_sequence)\n new_max_b = np.max(conv_new)\n\n curr_sequence = new_sequence\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 # Random perturbation to avoid local minima\n if iteration_count % perturbation_freq == 0: # Increased frequency\n perturbation = np.random.normal(0, 0.05, 1000)\n perturbation = np.clip(perturbation, -0.1, 0.1)\n curr_sequence += perturbation\n curr_sequence = np.maximum(0.0, curr_sequence)\n curr_sequence = curr_sequence / np.sum(curr_sequence) * initial_sum\n # Ensure the sum is maintained to avoid evaluate_sequence issues\n if np.sum(curr_sequence) < 0.001:\n curr_sequence = curr_sequence * (0.001 / np.sum(curr_sequence))\n\n # Early stopping if no improvement for many iterations\n if iteration_count % 100 == 0:\n print(f\"Iteration {iteration_count}, current best score: {best_score}\")\n if no_improvement_counter > 200: # Increased threshold\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 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": 489.5367934964597,
65 "env/all/time/policy/min": 202.45415496826172,
66 "env/all/time/policy/max": 683.1813368797302,
67 "env/all/time/env_step": 2623.7749838302843,
68 "env/all/time/env_step/min": 0.48737096786499023,
69 "env/all/time/env_step/max": 4826.177067518234,
70 "env/all/time/reward_compute": 3.632158041000366e-07,
71 "env/all/time/reward_compute/min": 2.0489096641540527e-07,
72 "env/all/time/reward_compute/max": 8.344650268554688e-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": 0.02489420957863331,
77 "advantage/min": -1.0,
78 "advantage/max": 24.462724685668945,
79 "time/assemble_training_data": 9.109660387039185,
80 "time/kl_vs_base": 135.62308621406555,
81 "kl_policy_base": 0.0006951516261324286,
82 "time/train": 1058.1998116970062,
83 "time/save_checkpoint": 21.63419461250305,
84 "time/total": 6739.124696969986
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