Views
No views yet
ac1. Checkpoint saved
after training step 28 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 28,
3 "progress/batch": 28,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.58,
6 "puct/buffer_size": 456,
7 "puct/sampled_size": 8,
8 "puct/T": 14336,
9 "puct/scale_last": 0.5741309407896953,
10 "puct/buffer_value/mean": -1.5287879022563498,
11 "puct/buffer_value/std": 0.09248769031733592,
12 "puct/buffer_value/min": -2.0797874683482562,
13 "puct/buffer_value/max": -1.505656527558561,
14 "puct/buffer_timestep/mean": 13.24561403508772,
15 "puct/buffer_timestep/std": 8.22977201567446,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 27.0,
18 "puct/buffer_construction_len/mean": 1131.5570175438597,
19 "puct/buffer_construction_len/std": 770.1525260723789,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7850.0,
22 "puct/sampled_value/mean": -1.5056566504994908,
23 "puct/sampled_value/std": 5.068941096399276e-08,
24 "puct/sampled_value/min": -1.5056567210973044,
25 "puct/sampled_value/max": -1.505656527558561,
26 "puct/sampled_timestep/mean": 27.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 27.0,
29 "puct/sampled_timestep/max": 27.0,
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": 6245.0690150260925,
35 "env/all/ac_tokens_per_turn": 8505.60546875,
36 "env/all/ob_tokens_per_turn": 3211.0,
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": 4354870,
41 "env/all/total_ob_tokens": 1644032,
42 "env/all/time/sampling_mean": 304.4350447077304,
43 "env/all/time/sampling_max": 410.53390431404114,
44 "env/all/time/env_step_mean": 3025.0142517788336,
45 "env/all/time/env_step_max": 5851.669358968735,
46 "env/all/reward/mean": 0.6042350399228245,
47 "env/all/reward/max": 0.6641620947772804,
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": 0.6042350399228245,
53 "env/all/correctness": 0.927734375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 1.5930809064674745,
57 "env/all/raw_score/min": 1.5056565275585598,
58 "env/all/raw_score/max": 26.414619169026018,
59 "env/all/initial_raw_score": -1.5056566504994908,
60 "env/all/initial_raw_score/min": -1.5056567210973044,
61 "env/all/initial_raw_score/max": -1.505656527558561,
62 "env/all/msg": "Success; raw_score=1.5056565275585603",
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 This approach uses an advanced LP-based optimization strategy focusing on convolution constraints,\n with increased exploration through perturbations and dynamic sequence expansion.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Start from known best construction if it exists\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 initial sequences with structural patterns\n initial_sequences = []\n for _ in range(100): # Increased number of initial sequences\n base = np.zeros(512)\n if np.random.rand() < 0.05: # Gauss-like decay\n base = np.exp(-np.linspace(0, 1, 512)**2) * 0.1 + 0.02\n elif np.random.rand() < 0.05: # Uniform\n base = np.random.uniform(0.01, 0.1, size=512)\n elif np.random.rand() < 0.05: # Exponential decay\n base = np.random.exponential(scale=0.1, size=512) * 0.05\n elif np.random.rand() < 0.05: # Periodic\n base = np.sin(np.linspace(0, 2 * np.pi, 512)) * 0.05 + 0.05\n elif np.random.rand() < 0.05: # Sine wave with varying amplitude\n base = np.sin(np.linspace(0, 2 * np.pi, 512)) * 0.1 + 0.05\n else: # Random\n base = np.random.rand(512)\n sum_base = np.sum(base)\n scale_factor = 0.5 / (sum_base if sum_base > 0.0 else 0.5)\n initial_sequences.append([max(0.0, x * scale_factor) for x in base])\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\n def perturb_sequence(seq):\n \"\"\"Perturbs elements based on convolution peaks for targeted improvement.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n conv = np.convolve(seq, seq)\n top_indices = np.argsort(conv)[-100:] # Top 100 positions with highest convolution\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 if np.random.rand() < 0.2: # Occasionally perturb a random element\n random_idx = np.random.randint(0, len(seq))\n indices.append(random_idx)\n scale = 0.2 * np.std(seq) # Slightly larger perturbation\n for idx in indices:\n new_val = max(0.0, seq[idx] + np.random.normal(0, scale))\n seq = [new_val if i == idx else seq[i] for i in range(len(seq))]\n return seq\n\n def random_walk(seq):\n \"\"\"Adds small random perturbations to all elements.\"\"\"\n scale = 0.1 * np.std(seq)\n new_seq = [max(0.0, x + np.random.normal(0, scale)) for x in seq]\n return new_seq\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using dynamic 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 = 0.9 * max_b_val if max_b_val > 0 else 0.1\n tight_positions = np.where(conv >= threshold)[0]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n g_fun = solve_convolution_lp(sequence, max_b_val, tight_positions)\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 # Golden-section search for optimal t\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(100):\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):\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) <= rhs)\n problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.GLPK, verbose=False, eps=1e-6)\n return g.value if problem.status == cp.OPTIMAL else None\n except:\n return None\n\n while time.time() < deadline:\n # Try 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 # Expand sequence if possible\n if len(current_sequence) < 1500:\n n = len(current_sequence)\n new_val = max(0.0, np.max(current_sequence) * 0.5)\n new_seq = current_sequence + [new_val]\n new_sum = np.sum(new_seq)\n if new_sum < 0.01:\n scale_factor = 0.01 / new_sum\n new_seq = [x * scale_factor for x in new_seq]\n current_sequence = new_seq\n else:\n current_sequence = perturb_sequence(current_sequence)\n\n # Occasionally try random walk for exploration\n if np.random.rand() < 0.05:\n current_sequence = random_walk(current_sequence)\n\n # Add structured perturbation strategy\n if np.random.rand() < 0.03:\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)[-10:] # Focus on most impactful positions\n peak_indices = np.array([min(idx, n-1) for idx in peak_indices])\n # Try to balance the values around the peak\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] * 1.2)\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": 304.4350447077304,
65 "env/all/time/policy/min": 120.72575569152832,
66 "env/all/time/policy/max": 410.53390431404114,
67 "env/all/time/env_step": 3025.0142517788336,
68 "env/all/time/env_step/min": 0.8736052513122559,
69 "env/all/time/env_step/max": 5851.669358968735,
70 "env/all/time/reward_compute": 4.805624485015869e-07,
71 "env/all/time/reward_compute/min": 2.7939677238464355e-07,
72 "env/all/time/reward_compute/max": 1.2442469596862793e-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": 0.00782079715281725,
77 "advantage/min": -1.0,
78 "advantage/max": 0.8529410362243652,
79 "time/assemble_training_data": 10.290679693222046,
80 "time/kl_vs_base": 91.39946269989014,
81 "kl_policy_base": 0.0010929028503596783,
82 "time/train": 562.7021291255951,
83 "time/save_checkpoint": 20.88383984565735,
84 "time/total": 6932.638441801071
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