Views
No views yet
ac1. Checkpoint saved
after training step 5 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 5,
3 "progress/batch": 5,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.12,
6 "puct/buffer_size": 88,
7 "puct/sampled_size": 8,
8 "puct/T": 2560,
9 "puct/scale_last": 0.39823737686334404,
10 "puct/buffer_value/mean": -1.5787270606088257,
11 "puct/buffer_value/std": 0.1542856379345064,
12 "puct/buffer_value/min": -2.000000000000008,
13 "puct/buffer_value/max": -1.5073856639307264,
14 "puct/buffer_timestep/mean": 1.7272727272727273,
15 "puct/buffer_timestep/std": 1.6006197146962735,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 4.0,
18 "puct/buffer_construction_len/mean": 1494.6704545454545,
19 "puct/buffer_construction_len/std": 1356.2901080922488,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7397.0,
22 "puct/sampled_value/mean": -1.5089180594225788,
23 "puct/sampled_value/std": 0.0009739046589887272,
24 "puct/sampled_value/min": -1.510074987988364,
25 "puct/sampled_value/max": -1.5073856639307264,
26 "puct/sampled_timestep/mean": 4.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 4.0,
29 "puct/sampled_timestep/max": 4.0,
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": 4368.445328474045,
35 "env/all/ac_tokens_per_turn": 8688.45703125,
36 "env/all/ob_tokens_per_turn": 2899.0,
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": 4448490,
41 "env/all/total_ob_tokens": 1484288,
42 "env/all/time/sampling_mean": 309.73549054702744,
43 "env/all/time/sampling_max": 398.9806203842163,
44 "env/all/time/env_step_mean": 1755.5050143566914,
45 "env/all/time/env_step_max": 3973.2817685604095,
46 "env/all/reward/mean": 0.4392493215543265,
47 "env/all/reward/max": 0.6634495668222538,
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": 0.4392493215543265,
53 "env/all/correctness": 0.693359375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 2.182813704751918,
57 "env/all/raw_score/min": 1.5072735643724018,
58 "env/all/raw_score/max": 126.15830955246162,
59 "env/all/initial_raw_score": -1.5089180594225793,
60 "env/all/initial_raw_score/min": -1.510074987988364,
61 "env/all/initial_raw_score/max": -1.5073856639307264,
62 "env/all/msg": "Success; raw_score=2.0241788830382457",
63 "env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy import optimize\nfrom scipy.optimize import minimize_scalar\n\nlinprog = optimize.linprog\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Proposes an optimal sequence of non-negative numbers that minimizes the evaluation score.\n The algorithm uses a combination of random initialization, constrained LP optimization,\n targeted perturbations, and adaptive line search to improve sequences incrementally.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Start with a random initialization for better exploration\n n = 1000\n best_sequence = [np.random.uniform(0.0, 1.0) for _ in range(n)]\n best_sequence = [float(x) for x in best_sequence]\n best_sequence = [max(0.0, x) for x in best_sequence]\n curr_sequence = best_sequence.copy()\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 # Fallback to targeted perturbations\n idx = perturb_targeted(curr_sequence)\n curr_sequence[idx] = max(0.0, curr_sequence[idx] + np.random.randn() * 1.5)\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 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 constrained LP optimization with reduced constraints.\"\"\"\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 the sequence 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 only the constraint for the current max convolution position\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 # Adaptive line search over t to minimize the evaluation function\n def objective(t):\n new_sequence = [(1 - t) * x + t * y for x, y in zip(sequence, normalized_g_fun)]\n return evaluate_sequence(new_sequence)\n \n # Perform a line search to find the optimal t\n result = minimize_scalar(objective, bounds=(0.0, 1.0), method='bounded')\n if result.success:\n best_t = result.x\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, compute the convolution to find the max b position\n conv = np.convolve(normalized_sequence, normalized_sequence)\n max_conv_idx = np.argmax(conv)\n max_conv_val = conv[max_conv_idx]\n # Identify elements contributing to this maximum\n perturb_elements = set()\n for i in range(n):\n j = max_conv_idx - i\n if 0 <= j < n:\n perturb_elements.add(i)\n perturb_elements.add(j)\n if perturb_elements:\n idx = perturb_targeted(normalized_sequence, max_conv_idx)\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 else:\n # If no elements are contributing, fall back to random\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() * 1.0)\n return new_seq\n\ndef perturb_targeted(sequence, max_conv_idx):\n \"\"\"Perturb elements contributing to the maximum convolution position.\"\"\"\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 solve_convolution_lp(f_sequence, rhs):\n \"\"\"Solves LP to maximize sum(b) s.t. conv(f, b) <= rhs, b >= 0 with reduced constraints.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n conv = np.convolve(f_sequence, f_sequence)\n max_conv_idx = np.argmax(conv)\n\n # Only use the constraint at the position with maximum convolution\n c = -np.ones(n)\n a_ub = []\n b_ub = []\n\n # Add the constraint for the max_conv_idx\n row = np.zeros(n)\n for i in range(n):\n j = max_conv_idx - 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 # Solve the LP with optimized settings\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 to focus on quality over speed\n \"disp\": False,\n },\n )\n\n if result.success:\n return result.x\n return None\n```",
64 "env/all/time/policy": 309.73549054702744,
65 "env/all/time/policy/min": 103.66378784179688,
66 "env/all/time/policy/max": 398.9806203842163,
67 "env/all/time/env_step": 1755.5050143566914,
68 "env/all/time/env_step/min": 0.005817890167236328,
69 "env/all/time/env_step/max": 3973.2817685604095,
70 "env/all/time/reward_compute": 2.789311110973358e-07,
71 "env/all/time/reward_compute/min": 1.8998980522155762e-07,
72 "env/all/time/reward_compute/max": 3.725290298461914e-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": 0.021706420928239822,
77 "advantage/min": -1.0,
78 "advantage/max": 7.909782409667969,
79 "time/assemble_training_data": 5.544961452484131,
80 "time/kl_vs_base": 83.20489954948425,
81 "kl_policy_base": 0.0006196670001372695,
82 "time/train": 553.8139955997467,
83 "time/save_checkpoint": 34.81086206436157,
84 "time/total": 5047.45773434639
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