Views
No views yet
ac1. Checkpoint saved
after training step 8 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 8,
3 "progress/batch": 8,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.18,
6 "puct/buffer_size": 136,
7 "puct/sampled_size": 8,
8 "puct/T": 4096,
9 "puct/scale_last": 0.3984351123685592,
10 "puct/buffer_value/mean": -1.5537727372910435,
11 "puct/buffer_value/std": 0.12862603608804507,
12 "puct/buffer_value/min": -2.000000000000008,
13 "puct/buffer_value/max": -1.5071879284255112,
14 "puct/buffer_timestep/mean": 3.235294117647059,
15 "puct/buffer_timestep/std": 2.462170533700747,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 7.0,
18 "puct/buffer_construction_len/mean": 1320.0808823529412,
19 "puct/buffer_construction_len/std": 1116.316388943005,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7397.0,
22 "puct/sampled_value/mean": -1.507234450578953,
23 "puct/sampled_value/std": 3.8622147641286394e-05,
24 "puct/sampled_value/min": -1.5072726535996397,
25 "puct/sampled_value/max": -1.5071879284255112,
26 "puct/sampled_timestep/mean": 7.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 7.0,
29 "puct/sampled_timestep/max": 7.0,
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": 5324.808835506439,
35 "env/all/ac_tokens_per_turn": 8855.916015625,
36 "env/all/ob_tokens_per_turn": 3454.125,
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": 4534229,
41 "env/all/total_ob_tokens": 1768512,
42 "env/all/time/sampling_mean": 334.59463317878544,
43 "env/all/time/sampling_max": 427.9699902534485,
44 "env/all/time/env_step_mean": 2586.8037245539017,
45 "env/all/time/env_step_max": 4905.15492939949,
46 "env/all/reward/mean": 0.5715380269604036,
47 "env/all/reward/max": 0.6634877924583534,
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": 0.5715380269604036,
53 "env/all/correctness": 0.873046875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 1.5514121115555886,
57 "env/all/raw_score/min": 1.5071867255611808,
58 "env/all/raw_score/max": 10.49046029556176,
59 "env/all/initial_raw_score": -1.5072344505789526,
60 "env/all/initial_raw_score/min": -1.5072726535996397,
61 "env/all/initial_raw_score/max": -1.5071879284255112,
62 "env/all/msg": "Success; raw_score=1.5071879284255112",
63 "env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy import optimize\nimport random\n\nlinprog = optimize.linprog\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers that minimizes the evaluation score \n using a hybrid of linear programming with adaptive line search and targeted perturbations.\n Focuses on improving initial sequences and enhancing line search for better convergence.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Use the last best construction if available\n best_sequence = None\n if 'height_sequence_1' in globals() and isinstance(height_sequence_1, (list, np.ndarray)):\n best_sequence = list(height_sequence_1)\n else:\n # Generate a better initial sequence: random uniform sequence normalized\n n = 1000\n random_seq = np.random.uniform(0.0, 1.0, n)\n random_seq *= np.sqrt(2 * n) / np.sum(random_seq)\n best_sequence = random_seq.tolist()\n \n curr_sequence = [float(x) for x in best_sequence]\n best_score = evaluate_sequence(curr_sequence)\n\n while time.time() < deadline:\n try:\n h_function = get_good_direction_to_move_into(curr_sequence)\n if h_function is not None:\n curr_sequence = h_function\n else:\n # Apply aggressive targeted perturbation\n idx = perturb_targeted(curr_sequence)\n # Smaller perturbation to avoid overshooting\n curr_sequence = perturb_sequence(curr_sequence, idx, 5.0)\n except Exception:\n print(\"Evaluation error, 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 # Final cleanup to ensure non-negative values\n return [float(max(0.0, x)) for x in best_sequence]\n\ndef get_good_direction_to_move_into(sequence):\n \"\"\"Returns a better sequence using LP with enhanced line search\"\"\"\n n = len(sequence)\n if n == 0:\n return None\n\n sum_sequence = np.sum(sequence)\n if sum_sequence <= 0.0:\n return None\n\n # Normalize for LP\n normalized_sequence = [x * np.sqrt(2 * n) / sum_sequence for x in sequence]\n rhs = np.max(np.convolve(normalized_sequence, normalized_sequence))\n\n # Solve LP with increased time limit\n g_fun = solve_convolution_lp(normalized_sequence, rhs)\n if g_fun is not None and np.sum(g_fun) > 0.0:\n sum_g = np.sum(g_fun)\n normalized_g_fun = [x * np.sqrt(2 * n) / sum_g for x in g_fun]\n # Perform more refined line search with increased t_candidate points\n best_t = 0.5\n best_new_score = np.inf\n t_values = np.linspace(0.01, 0.99, 100) # Reduced number of points for speed\n for t_candidate in t_values:\n new_sequence = [(1 - t_candidate) * x + t_candidate * y for x, y in zip(sequence, normalized_g_fun)]\n try:\n new_score = evaluate_sequence(new_sequence)\n if new_score < best_new_score:\n best_new_score = new_score\n best_t = t_candidate\n except:\n pass\n new_sequence = [(1 - best_t) * x + best_t * y for x, y in zip(sequence, normalized_g_fun)]\n return new_sequence\n \n # If LP fails, find max_conv_idx and target elements contributing to it\n conv = np.convolve(normalized_sequence, normalized_sequence)\n max_conv_idx = np.argmax(conv)\n contributions = []\n for i in range(n):\n j = max_conv_idx - i\n if 0 <= j < n:\n contributions.append((i, abs(normalized_sequence[i] * normalized_sequence[j])))\n if contributions:\n sorted_contributions = sorted(contributions, key=lambda x: x[1], reverse=True)\n idx = sorted_contributions[0][0]\n new_seq = normalized_sequence.copy()\n new_seq[idx] = max(0.0, new_seq[idx] - np.random.randn() * 3.0)\n return new_seq\n else:\n idx = np.random.randint(0, len(normalized_sequence))\n new_seq = normalized_sequence.copy()\n new_seq[idx] = max(0.0, new_seq[idx] - np.random.randn() * 2.0)\n return new_seq\n\ndef perturb_targeted(sequence, max_conv_idx):\n \"\"\"Targeted perturbation of elements contributing to max_conv_idx\"\"\"\n n = len(sequence)\n contributions = []\n for i in range(n):\n j = max_conv_idx - i\n if 0 <= j < n:\n contributions.append((i, abs(sequence[i] * sequence[j])))\n if contributions:\n sorted_contributions = sorted(contributions, key=lambda x: x[1], reverse=True)\n return sorted_contributions[0][0]\n else:\n return np.random.randint(0, n)\n\ndef perturb_sequence(sequence, idx, perturb_amount):\n \"\"\"Perturb the sequence at given index with smaller amount\"\"\"\n new_seq = sequence.copy()\n new_seq[idx] = max(0.0, new_seq[idx] - np.random.randn() * perturb_amount)\n new_seq[idx] = min(1000.0, new_seq[idx])\n return new_seq\n\ndef solve_convolution_lp(f_sequence, rhs):\n \"\"\"Solves LP to maximize sum(b) s.t. conv(f, b) <= rhs, b >= 0\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n c = -np.ones(n)\n a_ub = []\n b_ub = []\n\n for k in range(2 * n - 1):\n row = np.zeros(n)\n for i in range(n):\n j = k - i\n if 0 <= j < n:\n row[j] = f_sequence[i]\n a_ub.append(row)\n b_ub.append(rhs)\n\n # Non-negativity constraints\n a_ub_nonneg = -np.eye(n)\n b_ub_nonneg = np.zeros(n)\n a_ub = np.vstack([a_ub, a_ub_nonneg])\n b_ub = np.hstack([b_ub, b_ub_nonneg])\n\n # Use resilient LP solver with increased time limit\n result = 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\": 30.0, # Reduced time limit for speed\n \"disp\": False,\n },\n )\n\n if result.success:\n return result.x\n return None\n```",
64 "env/all/time/policy": 334.59463317878544,
65 "env/all/time/policy/min": 137.5316503047943,
66 "env/all/time/policy/max": 427.9699902534485,
67 "env/all/time/env_step": 2586.8037245539017,
68 "env/all/time/env_step/min": 0.006443977355957031,
69 "env/all/time/env_step/max": 4905.15492939949,
70 "env/all/time/reward_compute": 5.029141902923584e-07,
71 "env/all/time/reward_compute/min": 2.0116567611694336e-07,
72 "env/all/time/reward_compute/max": 1.7397105693817139e-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": 0.017515182495117188,
77 "advantage/min": -1.0,
78 "advantage/max": 19.10958480834961,
79 "time/assemble_training_data": 6.131303787231445,
80 "time/kl_vs_base": 92.1794924736023,
81 "kl_policy_base": 0.000667551823426038,
82 "time/train": 593.5688090324402,
83 "time/save_checkpoint": 36.21559262275696,
84 "time/total": 6054.498106718063
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