Views
No views yet
ac1. Checkpoint saved
after training step 45 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 45,
3 "progress/batch": 45,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.92,
6 "puct/buffer_size": 728,
7 "puct/sampled_size": 8,
8 "puct/T": 23040,
9 "puct/scale_last": 0.5750370987848814,
10 "puct/buffer_value/mean": -1.519938993289323,
11 "puct/buffer_value/std": 0.07408994358679578,
12 "puct/buffer_value/min": -2.0797874683482562,
13 "puct/buffer_value/max": -1.5047503695633748,
14 "puct/buffer_timestep/mean": 21.747252747252748,
15 "puct/buffer_timestep/std": 13.136304368147167,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 44.0,
18 "puct/buffer_construction_len/mean": 1082.4038461538462,
19 "puct/buffer_construction_len/std": 612.8413147121971,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7850.0,
22 "puct/sampled_value/mean": -1.504751678797381,
23 "puct/sampled_value/std": 5.082036777028696e-07,
24 "puct/sampled_value/min": -1.5047519414347188,
25 "puct/sampled_value/max": -1.5047503695633748,
26 "puct/sampled_timestep/mean": 44.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 44.0,
29 "puct/sampled_timestep/max": 44.0,
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": 6249.78152179718,
35 "env/all/ac_tokens_per_turn": 7765.529296875,
36 "env/all/ob_tokens_per_turn": 3890.5,
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": 3975951,
41 "env/all/total_ob_tokens": 1991936,
42 "env/all/time/sampling_mean": 281.72924899403006,
43 "env/all/time/sampling_max": 391.18807077407837,
44 "env/all/time/env_step_mean": 3084.0212873602286,
45 "env/all/time/env_step_max": 5876.835255146027,
46 "env/all/reward/mean": 0.6172289149208512,
47 "env/all/reward/max": 0.6645621457587517,
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": 0.6172289149208512,
53 "env/all/correctness": 0.94140625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 1.5512514451067865,
57 "env/all/raw_score/min": 1.5047501572823514,
58 "env/all/raw_score/max": 8.164382133850847,
59 "env/all/initial_raw_score": -1.5047516787973811,
60 "env/all/initial_raw_score/min": -1.5047519414347188,
61 "env/all/initial_raw_score/max": -1.5047503695633748,
62 "env/all/msg": "Success; raw_score=1.5047501572823514",
63 "env/all/parsed_code": "```python\nimport time\nimport numpy as np\nimport cvxpy as cp\nimport copy\nimport random\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Optimizes a sequence of non-negative coefficients to minimize the evaluation function.\n Implements enhanced LP-based directional moves, dynamic threshold scaling, aggressive exploration, and refined expansion strategies.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 30 # Allow 30 seconds for final stabilization\n\n # Initialize from previous best if available\n prev = globals().get(\"height_sequence_1\", None)\n if prev is not None and isinstance(prev, (list, np.ndarray)) and len(prev) > 0:\n best_sequence = list(np.asarray(prev, dtype=float))\n else:\n initial_sequences = []\n for _ in range(150): # More diversity with 150 initial sequences\n base_len = min(2048, np.random.choice([256, 512, 1024, 2048]))\n base = np.zeros(base_len)\n \n # Generate more varied patterns\n if np.random.rand() < 0.25:\n # Staggered sinusoidal patterns with more control\n freqs = np.random.randint(1, 5, size=3)\n for f in freqs:\n base += np.sin(f * np.linspace(0, 2 * np.pi, base_len)) * 0.2\n elif np.random.rand() < 0.2:\n # Random exponential pulses with finer granularity\n for _ in range(3):\n pos = np.random.randint(0, base_len)\n base += np.exp(-((np.linspace(0, 1, base_len) - pos)**2)*10) * 0.2\n elif np.random.rand() < 0.2:\n # Random pulse trains with more structured spacing\n num_pulses = np.random.randint(3, 6)\n for _ in range(num_pulses):\n pos = np.random.randint(0, base_len)\n width = np.random.randint(1, 5)\n base[pos:pos+width] += np.random.rand() * 0.2\n elif np.random.rand() < 0.2:\n # Geometric decay sequences with different parameters\n r = np.random.uniform(0.6, 0.9)\n for i in range(base_len):\n base[i] = 1.0 * r**i\n base = [min(1000.0, max(0.0, x)) for x in base]\n else:\n # Default structured decay with increased frequency\n base = np.random.rand(base_len) * 0.5 + np.exp(-np.linspace(0, 1, base_len)**2 * 20) * 0.1\n \n sum_base = np.sum(base)\n scale_factor = np.sqrt(2 * base_len) / sum_base if sum_base > 0 else 0.5\n seq = [max(0.0, min(1000.0, x * scale_factor)) for x in base]\n initial_sequences.append(seq)\n best_sequence = initial_sequences[0].copy()\n for seq in initial_sequences:\n curr_score = evaluate_sequence(seq)\n best_score = evaluate_sequence(best_sequence)\n if curr_score < best_score:\n best_sequence = seq.copy()\n\n current_sequence = best_sequence.copy()\n best_score = float('inf')\n best_sequence = list(current_sequence)\n\n def perturb_sequence(seq):\n \"\"\"Strategic perturbation with increased exploration and tighter scaling.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n conv = np.convolve(seq, seq)\n top_indices = np.argsort(conv)[::-1] # Top indices\n indices = np.random.choice(top_indices, size=min(100, len(top_indices)), replace=False)\n indices = [min(idx, len(seq)-1) for idx in indices]\n scale = 0.7 * np.std(seq) # Larger perturbations\n new_seq = seq.copy()\n for idx in indices:\n if new_seq[idx] > 0.0:\n perturbation = np.random.normal(0, scale) * 0.5\n new_val = new_seq[idx] - perturbation\n adjustment = new_val - new_seq[idx]\n other_idx = np.random.randint(0, n)\n new_seq[other_idx] += adjustment\n new_seq[idx] = new_val\n new_seq = [min(1000.0, max(0.0, x)) for x in new_seq]\n return new_seq\n\n def random_walk(seq):\n \"\"\"Random walk with dynamic perturbation scale and adaptive target sum.\"\"\"\n n = len(seq)\n target_sum = np.sqrt(2 * n) # Optimal normalization\n current_sum = np.sum(seq)\n scale = 0.05 * np.std(seq) * np.random.uniform(0.6, 1.4)\n perturbation = np.random.uniform(-scale, scale, size=n)\n new_seq = [seq[i] + perturbation[i] for i in range(n)]\n adjustment = target_sum - np.sum(new_seq)\n new_seq[-1] += adjustment\n new_seq = [min(1000.0, max(0.0, x)) for x in new_seq]\n return new_seq\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Optimizes through adaptive constraint selection with dynamic thresholds.\"\"\"\n n = len(sequence)\n sum_sequence = np.sum(sequence)\n if sum_sequence <= 0.0:\n return None\n conv = np.convolve(sequence, sequence)\n max_b_val = np.max(conv)\n threshold = max(0.001 * max_b_val, 0.002 * max_b_val * np.sqrt(n)) if max_b_val > 0 else 0.05\n tight_positions = np.where(conv >= threshold)[0]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n \n # Use more aggressive threshold scaling for tighter constraints\n g_fun = solve_convolution_lp(sequence, threshold * 0.9, tight_positions, threshold_scale=0.01)\n if g_fun is None:\n return None\n sum_g = np.sum(g_fun)\n if sum_g <= 0.0:\n return None\n # Apply golden-section search with adaptive iterations\n def objective(t):\n new_seq = [ (1 - t) * x + t * y for x, y in zip(sequence, g_fun) ]\n return evaluate_sequence(new_seq)\n low, high = 0.0, 1.0\n for _ in range(150): # Increased iteration count\n t1 = low + (high - low) / 3\n t2 = high - (high - low) / 3\n f1 = objective(t1)\n f2 = objective(t2)\n if f1 < f2:\n high = t2\n else:\n low = t1\n best_t = (low + high) / 2\n new_seq = [ (1 - best_t) * x + best_t * y for x, y in zip(sequence, g_fun) ]\n return new_seq\n\n def solve_convolution_lp(f_sequence, rhs, tight_positions, threshold_scale=0.01):\n \"\"\"Solve with adaptive threshold_scale for tighter constraints.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n g = cp.Variable(n, nonneg=True)\n objective = cp.Minimize(-cp.sum(g)) # Maximize sum(g)\n constraints = []\n for k in tight_positions:\n coeff = np.zeros(n)\n for j in range(n):\n i = k - j\n if 0 <= i < n:\n coeff[j] = f_sequence[i]\n # Use tighter and dynamic scaling for constraints\n current_max_b = np.max(np.convolve(f_sequence, f_sequence))\n dynamic_threshold = 0.01 * current_max_b\n constraints.append(cp.sum(coeff * g) <= dynamic_threshold)\n problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.GLPK, verbose=False, eps=1e-6)\n if problem.status == cp.OPTIMAL:\n g_fun = g.value\n target_sum = np.sqrt(2 * n)\n current_sum = np.sum(g_fun)\n if current_sum > 0:\n g_fun = [x * (target_sum / current_sum) for x in g_fun]\n g_fun = [min(1000.0, max(0.0, x)) for x in g_fun]\n return g_fun\n else:\n return None\n except:\n return None\n\n def expand_sequence(seq):\n \"\"\"Expand with careful addition of new elements using balanced distribution.\"\"\"\n n = len(seq)\n if n >= 2000:\n return seq\n new_length = n + 1\n target_sum = np.sqrt(2 * new_length)\n current_sum = np.sum(seq)\n # Find the position with the highest convolution value\n conv = np.convolve(seq, seq)\n peak_pos = np.argmax(conv)\n peak_pos = min(peak_pos, n) # Ensure it's within bounds\n # Spread out the new value instead of concentrating it\n new_val = current_sum * 0.7 / new_length * np.random.uniform(0.9, 1.1)\n new_seq = np.zeros(new_length)\n new_seq[:peak_pos] = seq[:peak_pos]\n new_seq[peak_pos] = new_val\n new_seq[peak_pos+1:] = seq[peak_pos:]\n return [max(0.0, min(1000.0, x)) for x in new_seq]\n\n while time.time() < deadline:\n # LP-based move with adjusted threshold_scale\n h_function = get_good_direction_to_move_into(current_sequence)\n if h_function is not None:\n current_sequence = h_function\n else:\n if len(current_sequence) < 2000:\n current_sequence = expand_sequence(current_sequence)\n else:\n current_sequence = perturb_sequence(current_sequence)\n\n # Increased random walk frequency\n if np.random.rand() < 0.05:\n current_sequence = random_walk(current_sequence)\n\n # Limited targeted adjustment on convolution peaks\n if np.random.rand() < 0.01:\n seq = [max(0.0, x) for x in current_sequence]\n n = len(seq)\n conv = np.convolve(seq, seq)\n peak_indices = np.argsort(conv)[-30:] # Focus on top 30 peaks\n peak_indices = np.array([min(idx, n-1) for idx in peak_indices])\n for idx in peak_indices:\n if seq[idx] > 0.0:\n seq[idx] = max(0.0, seq[idx] * 0.8)\n for shift in [-1, 0, 1]:\n neighbor_idx = idx + shift\n if 0 <= neighbor_idx < n:\n seq[neighbor_idx] = min(1000.0, seq[neighbor_idx] * 1.1)\n current_sequence = [max(0.0, min(1000.0, x)) for x in seq]\n\n try:\n curr_score = evaluate_sequence(current_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = current_sequence.copy()\n print(f\"New best: {best_score}\")\n except Exception:\n pass\n\n return [float(max(0.0, x)) for x in best_sequence]\n```",
64 "env/all/time/policy": 281.72924899403006,
65 "env/all/time/policy/min": 127.25801706314087,
66 "env/all/time/policy/max": 391.18807077407837,
67 "env/all/time/env_step": 3084.0212873602286,
68 "env/all/time/env_step/min": 1.517451286315918,
69 "env/all/time/env_step/max": 5876.835255146027,
70 "env/all/time/reward_compute": 5.192123353481293e-07,
71 "env/all/time/reward_compute/min": 2.086162567138672e-07,
72 "env/all/time/reward_compute/max": 2.380460500717163e-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": 0.0047726621851325035,
77 "advantage/min": -1.0,
78 "advantage/max": 0.7588322162628174,
79 "time/assemble_training_data": 5.56314754486084,
80 "time/kl_vs_base": 96.2221372127533,
81 "kl_policy_base": 0.0008257756708189845,
82 "time/train": 557.107424736023,
83 "time/save_checkpoint": 24.295390844345093,
84 "time/total": 6935.555735588074
85}[2026-07-09T06:24:18+00:00] job=1812630 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T09:14:17+00:00] job=1813129 node=node-28 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:46:36+00:00] job=1813130 node=node-22 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-11T14:04:05+00:00] job=1824338 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-12T06:23:00+00:00] job=1827205 node=node-14 ngpu=6 ntrain=2 replicas=4 flash_attn=yes