Views
No views yet
ac1. Checkpoint saved
after training step 44 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 44,
3 "progress/batch": 44,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.9,
6 "puct/buffer_size": 712,
7 "puct/sampled_size": 8,
8 "puct/T": 22528,
9 "puct/scale_last": 0.5750355467356612,
10 "puct/buffer_value/mean": -1.5202802782879452,
11 "puct/buffer_value/std": 0.07488241187017297,
12 "puct/buffer_value/min": -2.0797874683482562,
13 "puct/buffer_value/max": -1.504751921612595,
14 "puct/buffer_timestep/mean": 21.247191011235955,
15 "puct/buffer_timestep/std": 12.847664850453647,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 43.0,
18 "puct/buffer_construction_len/mean": 1084.2556179775281,
19 "puct/buffer_construction_len/std": 619.5630159450008,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7850.0,
22 "puct/sampled_value/mean": -1.5047519419789774,
23 "puct/sampled_value/std": 8.103401660675995e-09,
24 "puct/sampled_value/min": -1.5047519496618476,
25 "puct/sampled_value/max": -1.504751921612595,
26 "puct/sampled_timestep/mean": 43.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 43.0,
29 "puct/sampled_timestep/max": 43.0,
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": 6253.853032588959,
35 "env/all/ac_tokens_per_turn": 7760.830078125,
36 "env/all/ob_tokens_per_turn": 4163.25,
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": 3973545,
41 "env/all/total_ob_tokens": 2131584,
42 "env/all/time/sampling_mean": 288.0968103031628,
43 "env/all/time/sampling_max": 365.8339774608612,
44 "env/all/time/env_step_mean": 3021.3760626530275,
45 "env/all/time/env_step_max": 5893.355804204941,
46 "env/all/reward/mean": 0.6179108204809829,
47 "env/all/reward/max": 0.6645620520063698,
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": 0.6179108204809829,
53 "env/all/correctness": 0.93359375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 1.5124439187749854,
57 "env/all/raw_score/min": 1.5047503695633748,
58 "env/all/raw_score/max": 2.1544529884866637,
59 "env/all/initial_raw_score": -1.5047519419789772,
60 "env/all/initial_raw_score/min": -1.5047519496618476,
61 "env/all/initial_raw_score/max": -1.504751921612595,
62 "env/all/msg": "Success; raw_score=1.5047515898833728",
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 Focuses on generating more diverse random sequences, aggressive reduction of convolution peaks,\n and improved LP-based steps for directional optimization.\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 # Generate diverse random initial sequences\n initial_sequences = []\n for _ in range(200): # Increase diversity with 200 initial sequences\n base_len = min(2048, np.random.choice([256, 512, 1024, 2048]))\n base = np.random.rand(base_len) * 0.5\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 targeting convolution peaks with dynamic scaling.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n conv = np.convolve(seq, seq)\n top_indices = np.argsort(conv)[-max(200, int(n * 0.05)):] # Top 200 positions\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 = 0.1 * np.std(seq)\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 threshold and structured normalization.\"\"\"\n n = len(seq)\n target_sum = np.sqrt(2 * n)\n current_sum = np.sum(seq)\n scale = 0.1 * np.std(seq) * np.random.uniform(0.6, 1.4)\n perturbation = np.random.uniform(-scale, scale, size=n) * 0.5\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 \"\"\"Optimized LP with adaptive threshold and constraint selection.\"\"\"\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(0.5 * max_b_val, 0.01 * 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 adaptive threshold scale for tighter constraints\n g_fun = solve_convolution_lp(sequence, threshold * max(0.3, 0.5 - np.random.rand()), tight_positions, threshold_scale=0.6)\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 \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 \n low, high = 0.0, 1.0\n for _ in range(150):\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=0.5):\n \"\"\"Solve with adaptive threshold_scale and constraint relaxation.\"\"\"\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))\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 constraints.append(cp.sum(coeff * g) <= threshold_scale * rhs)\n \n # Add a small constraint relaxation to avoid infeasibility\n constraints.append((cp.sum(g) >= 0.01 * np.sqrt(2 * n)))\n \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 based on existing structure.\"\"\"\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 # Optimize new_val using adaptive heuristic\n new_val = (target_sum - current_sum) * 0.5\n new_seq = seq + [max(0.0, min(1000.0, new_val))]\n return new_seq\n\n def contract_sequence(seq):\n \"\"\"Contract with structured removal of elements.\"\"\"\n n = len(seq)\n if n <= 1:\n return seq\n # Randomly remove one element\n remove_idx = np.random.randint(0, n)\n new_seq = seq[:remove_idx] + seq[remove_idx+1:]\n new_seq = [min(1000.0, max(0.0, x)) for x in new_seq]\n return new_seq\n\n # Add periodic restarts\n restart_counter = 0\n restart_interval = 100 # Restart every 100 iterations\n\n while time.time() < deadline:\n # LP-based move\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) < 1500:\n current_sequence = expand_sequence(current_sequence)\n else:\n current_sequence = perturb_sequence(current_sequence)\n\n # Random walk for exploration with structured perturbations\n if np.random.rand() < 0.05:\n current_sequence = random_walk(current_sequence)\n\n # Targeted adjustment on convolution peaks with diversified patterns\n if np.random.rand() < 0.02:\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)[-20:] # Focus on top 20 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] * 0.85)\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] * 1.05)\n current_sequence = [max(0.0, min(1000.0, x)) for x in seq]\n\n # Periodic restarts\n restart_counter += 1\n if restart_counter >= restart_interval:\n restart_counter = 0\n # Generate new initial sequence\n base_len = min(2048, np.random.choice([256, 512, 1024, 2048]))\n base = np.random.rand(base_len) * 0.5\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 current_sequence = 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": 288.0968103031628,
65 "env/all/time/policy/min": 126.9912486076355,
66 "env/all/time/policy/max": 365.8339774608612,
67 "env/all/time/env_step": 3021.3760626530275,
68 "env/all/time/env_step/min": 8.042646408081055,
69 "env/all/time/env_step/max": 5893.355804204941,
70 "env/all/time/reward_compute": 6.360933184623718e-07,
71 "env/all/time/reward_compute/min": 2.2351741790771484e-07,
72 "env/all/time/reward_compute/max": 1.944601535797119e-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": 0.004892850294709206,
77 "advantage/min": -1.0,
78 "advantage/max": 1.899369716644287,
79 "time/assemble_training_data": 6.033586740493774,
80 "time/kl_vs_base": 99.03146958351135,
81 "kl_policy_base": 0.000876234145835042,
82 "time/train": 573.2752017974854,
83 "time/save_checkpoint": 23.698670148849487,
84 "time/total": 6958.256349802017
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