Views
No views yet
ac1. Checkpoint saved
after training step 31 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 31,
3 "progress/batch": 31,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.64,
6 "puct/buffer_size": 504,
7 "puct/sampled_size": 8,
8 "puct/T": 15872,
9 "puct/scale_last": 0.4936516100901467,
10 "puct/buffer_value/mean": -1.521068968865627,
11 "puct/buffer_value/std": 0.07416355263427298,
12 "puct/buffer_value/min": -2.000000000000008,
13 "puct/buffer_value/max": -1.5063483899098538,
14 "puct/buffer_timestep/mean": 14.746031746031745,
15 "puct/buffer_timestep/std": 9.095556658040623,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 30.0,
18 "puct/buffer_construction_len/mean": 1086.3710317460318,
19 "puct/buffer_construction_len/std": 597.0358222499963,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7397.0,
22 "puct/sampled_value/mean": -1.5063602301188044,
23 "puct/sampled_value/std": 6.677419467850976e-08,
24 "puct/sampled_value/min": -1.5063602553666806,
25 "puct/sampled_value/max": -1.5063600534508914,
26 "puct/sampled_timestep/mean": 30.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 30.0,
29 "puct/sampled_timestep/max": 30.0,
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": 5358.966257572174,
35 "env/all/ac_tokens_per_turn": 8237.18359375,
36 "env/all/ob_tokens_per_turn": 3061.125,
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": 4217438,
41 "env/all/total_ob_tokens": 1567296,
42 "env/all/time/sampling_mean": 510.69254416553304,
43 "env/all/time/sampling_max": 695.9192101955414,
44 "env/all/time/env_step_mean": 2294.9849030189216,
45 "env/all/time/env_step_max": 4665.235037326813,
46 "env/all/reward/mean": 0.5955663535787399,
47 "env/all/reward/max": 0.6638545551934435,
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": 0.5955663535787399,
53 "env/all/correctness": 0.919921875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 1.5691514693602042,
57 "env/all/raw_score/min": 1.5063540432739218,
58 "env/all/raw_score/max": 9.774257920259718,
59 "env/all/initial_raw_score": -1.5063602301188042,
60 "env/all/initial_raw_score/min": -1.5063602553666806,
61 "env/all/initial_raw_score/max": -1.5063600534508914,
62 "env/all/msg": "Success; raw_score=1.5063600534508912",
63 "env/all/parsed_code": "```python\nimport time\nimport numpy as np\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 hybrid approach of genetic algorithm and line search optimization.\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 random uniform distribution\n n = 1000\n best_sequence = np.random.rand(n)\n best_sequence = best_sequence / np.sum(best_sequence) * np.sqrt(2 * n)\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 while time.time() < deadline:\n current_time = time.time()\n try:\n # Evaluate current sequence\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 top k indices with highest convolution values to form constraints\n k = 100 # Reduce the number of constraints to speed up LP\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(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)\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 # Normalize g_0 to have the same sum as curr_sequence\n g_0_normalized = g_0 / sum_g * sum_a\n\n # Perform line search to find optimal alpha\n def find_optimal_alpha(curr_seq, g0_norm):\n alpha_list = np.linspace(0, 1, 50) # Try 50 alpha values\n best_alpha = 0.5\n best_score = float('inf')\n for alpha in alpha_list:\n new_seq = (1 - alpha) * curr_seq + alpha * g0_norm\n new_seq = np.clip(new_seq, 0, 1000.0)\n sum_new = np.sum(new_seq)\n if sum_new < 0.01:\n continue\n new_seq_normalized = new_seq / sum_new * np.sqrt(2 * len(new_seq))\n try:\n new_score = evaluate_sequence(new_seq_normalized.tolist())\n except:\n new_score = float('inf')\n if new_score < best_score:\n best_score = new_score\n best_alpha = alpha\n return best_alpha\n\n optimal_alpha = find_optimal_alpha(curr_sequence, g_0_normalized)\n new_sequence = (1 - optimal_alpha) * curr_sequence + optimal_alpha * g_0_normalized\n new_sequence = np.clip(new_sequence, 0, 1000.0)\n new_sequence = new_sequence / np.sum(new_sequence) * np.sqrt(2 * 1000)\n\n # Evaluate and update best sequence\n curr_score = evaluate_sequence(new_sequence.tolist())\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = new_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 % 100 == 0:\n perturbation = np.random.normal(0, 0.01, 1000)\n perturbation = np.clip(perturbation, -0.05, 0.05)\n curr_sequence += perturbation\n curr_sequence = np.maximum(0.0, curr_sequence)\n curr_sequence = curr_sequence / np.sum(curr_sequence) * np.sqrt(2 * 1000)\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: # 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": 510.69254416553304,
65 "env/all/time/policy/min": 235.51084566116333,
66 "env/all/time/policy/max": 695.9192101955414,
67 "env/all/time/env_step": 2294.9849030189216,
68 "env/all/time/env_step/min": 0.5287649631500244,
69 "env/all/time/env_step/max": 4665.235037326813,
70 "env/all/time/reward_compute": 4.5029446482658386e-07,
71 "env/all/time/reward_compute/min": 1.601874828338623e-07,
72 "env/all/time/reward_compute/max": 1.084059476852417e-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": 0.012758512049913406,
77 "advantage/min": -1.0,
78 "advantage/max": 18.11575698852539,
79 "time/assemble_training_data": 6.637221574783325,
80 "time/kl_vs_base": 136.17140126228333,
81 "kl_policy_base": 0.0007440380868501961,
82 "time/train": 1073.6179466247559,
83 "time/save_checkpoint": 21.10323739051819,
84 "time/total": 6601.8775091171265
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