Views
No views yet
ac1. Checkpoint saved
after training step 1 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 1,
3 "progress/batch": 1,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.04,
6 "puct/buffer_size": 24,
7 "puct/sampled_size": 8,
8 "puct/T": 512,
9 "puct/scale_last": 0.3936487623820777,
10 "puct/buffer_value/mean": -1.7566130802550568,
11 "puct/buffer_value/std": 0.20913895210285907,
12 "puct/buffer_value/min": -2.000000000000008,
13 "puct/buffer_value/max": -1.5119742784119927,
14 "puct/buffer_timestep/mean": -0.3333333333333333,
15 "puct/buffer_timestep/std": 0.4714045207910317,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 0.0,
18 "puct/buffer_construction_len/mean": 2612.6666666666665,
19 "puct/buffer_construction_len/std": 2229.7828681037286,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7397.0,
22 "puct/sampled_value/mean": -1.5233171741158662,
23 "puct/sampled_value/std": 0.011040902645112632,
24 "puct/sampled_value/min": -1.5444152228264785,
25 "puct/sampled_value/max": -1.5119742784119927,
26 "puct/sampled_timestep/mean": 0.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 0.0,
29 "puct/sampled_timestep/max": 0.0,
30 "puct/sampled_construction_len/mean": 1153.375,
31 "puct/sampled_construction_len/std": 179.3871076053126,
32 "puct/sampled_construction_len/min": 1000.0,
33 "puct/sampled_construction_len/max": 1493.0,
34 "time/sampling": 5311.143718242645,
35 "env/all/ac_tokens_per_turn": 7981.490234375,
36 "env/all/ob_tokens_per_turn": 2437.5,
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": 4086523,
41 "env/all/total_ob_tokens": 1248000,
42 "env/all/time/sampling_mean": 464.9060324449092,
43 "env/all/time/sampling_max": 654.2920558452606,
44 "env/all/time/env_step_mean": 2026.538373509422,
45 "env/all/time/env_step_max": 4668.935054779053,
46 "env/all/reward/mean": 0.4435371799934764,
47 "env/all/reward/max": 0.662511660560165,
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": 0.4435371799934764,
53 "env/all/correctness": 0.75390625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 10.279874669399897,
57 "env/all/raw_score/min": 1.5094073854177394,
58 "env/all/raw_score/max": 1999.9999999999995,
59 "env/all/initial_raw_score": -1.5233171741158662,
60 "env/all/initial_raw_score/min": -1.5444152228264785,
61 "env/all/initial_raw_score/max": -1.5119742784119927,
62 "env/all/msg": "RuntimeError: Program execution failed: UnboundLocalError: cannot access local variable 't' where it is not associated with a value",
63 "env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy import optimize\n\nlinprog = optimize.linprog\n\ndef get_good_direction_to_move_into(sequence):\n \"\"\"Returns a better sequence using LP optimization or perturbation.\"\"\"\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 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 # Attempt to find a better sequence via LP\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 old_score = evaluate_sequence(sequence)\n new_sequence = [(1 - t) * x + t * y for x, y in zip(sequence, normalized_g_fun)]\n new_score = evaluate_sequence(new_sequence)\n if new_score < old_score:\n t = 0.2 # Larger step if improvement is found\n else:\n t = 0.05\n new_sequence = [(1 - t) * x + t * y for x, y in zip(sequence, normalized_g_fun)]\n return new_sequence\n\n # If LP fails, apply random perturbation with cooling schedule\n idx = np.random.randint(0, len(sequence))\n curr_seq = sequence.copy()\n curr_seq[idx] = max(0.0, curr_seq[idx] + np.random.randn() * 0.5)\n return curr_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 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 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 # Increase LP time to allow deeper exploration\n result = linprog(\n c,\n A_ub=a_ub,\n b_ub=b_ub,\n options={\n \"time_limit\": 30.0, # Increased from 10 to 30 seconds\n \"disp\": False,\n },\n )\n if result.success:\n return result.x\n return None\n\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"Hybrid LP and adaptive gradient-based search with cooling perturbations.\"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Use the last best construction if available\n prev = globals().get(\"GLOBAL_BEST_CONSTRUCTION\")\n if np.random.rand() < 0.5 and isinstance(prev, (list, tuple, np.ndarray)) and len(prev) > 0:\n best_sequence = list(np.asarray(prev, dtype=float))\n else:\n # Generate a few initial sequences\n # Try a uniform sequence\n uniform_seq = np.ones(1000) / 1000.0\n # Try a random sequence\n random_seq = [float(np.random.randn()) for _ in range(1000)]\n random_seq = [max(0.0, x) for x in random_seq]\n sum_random = np.sum(random_seq)\n if sum_random < 0.01:\n random_seq = [x * 0.01 / sum_random for x in random_seq]\n best_sequence = random_seq\n\n curr_sequence = best_sequence.copy()\n best_score = evaluate_sequence(best_sequence)\n start_time = time.time()\n total_time_budget = deadline - start_time\n\n while time.time() < deadline:\n # Attempt to find a new direction\n h_function = get_good_direction_to_move_into(curr_sequence)\n if h_function is None:\n print(\"LP failed, using temperature-based cooling perturbation.\")\n remaining_time = deadline - time.time()\n # Adjust cooling schedule: larger steps initially\n step_size = 0.5 * (remaining_time / total_time_budget)\n idx = np.random.randint(0, len(curr_sequence))\n curr_sequence[idx] = max(0.0, curr_sequence[idx] + np.random.randn() * step_size)\n else:\n curr_sequence = h_function\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 return [float(max(0.0, x)) for x in best_sequence]\n```",
64 "env/all/time/policy": 464.9060324449092,
65 "env/all/time/policy/min": 208.35213327407837,
66 "env/all/time/policy/max": 654.2920558452606,
67 "env/all/time/env_step": 2026.538373509422,
68 "env/all/time/env_step/min": 0.004978179931640625,
69 "env/all/time/env_step/max": 4668.935054779053,
70 "env/all/time/reward_compute": 7.585622370243073e-07,
71 "env/all/time/reward_compute/min": 2.60770320892334e-07,
72 "env/all/time/reward_compute/max": 1.4454126358032227e-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": 0.025457359850406647,
77 "advantage/min": -1.0,
78 "advantage/max": 17.59844398498535,
79 "time/assemble_training_data": 8.385901689529419,
80 "time/kl_vs_base": 129.72728490829468,
81 "kl_policy_base": 0.0005377906491048634,
82 "time/train": 969.5151534080505,
83 "time/save_checkpoint": 10.334673166275024,
84 "time/total": 6430.796000957489
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