Views
No views yet
ac1. Checkpoint saved
after training step 21 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 21,
3 "progress/batch": 21,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.44,
6 "puct/buffer_size": 344,
7 "puct/sampled_size": 8,
8 "puct/T": 10752,
9 "puct/scale_last": 0.5741228667321363,
10 "puct/buffer_value/mean": -1.536316847765978,
11 "puct/buffer_value/std": 0.10539543742991385,
12 "puct/buffer_value/min": -2.0797874683482562,
13 "puct/buffer_value/max": -1.50566460161612,
14 "puct/buffer_timestep/mean": 9.744186046511627,
15 "puct/buffer_timestep/std": 6.2098684525475125,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 20.0,
18 "puct/buffer_construction_len/mean": 1174.389534883721,
19 "puct/buffer_construction_len/std": 882.4845966807882,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7850.0,
22 "puct/sampled_value/mean": -1.5056851014338681,
23 "puct/sampled_value/std": 1.5406978438183425e-05,
24 "puct/sampled_value/min": -1.505703506348427,
25 "puct/sampled_value/max": -1.50566460161612,
26 "puct/sampled_timestep/mean": 20.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 20.0,
29 "puct/sampled_timestep/max": 20.0,
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": 5712.875193119049,
35 "env/all/ac_tokens_per_turn": 7850.373046875,
36 "env/all/ob_tokens_per_turn": 3170.875,
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": 4019391,
41 "env/all/total_ob_tokens": 1623488,
42 "env/all/time/sampling_mean": 269.534353164956,
43 "env/all/time/sampling_max": 352.4924826622009,
44 "env/all/time/env_step_mean": 2843.082124957349,
45 "env/all/time/env_step_max": 5356.084791183472,
46 "env/all/reward/mean": 0.5281894585204324,
47 "env/all/reward/max": 0.6641598050711318,
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": 0.5281894585204324,
53 "env/all/correctness": 0.8359375,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 2.833340307873439,
57 "env/all/raw_score/min": 1.5056617183439782,
58 "env/all/raw_score/max": 494.8928096981323,
59 "env/all/initial_raw_score": -1.5056851014338681,
60 "env/all/initial_raw_score/min": -1.505703506348427,
61 "env/all/initial_raw_score/max": -1.50566460161612,
62 "env/all/msg": "Success; raw_score=1.505664552270154",
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 a combination of LP-based moves, targeted perturbations, and strategic expansion with improved parameters.\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 n = 500 # Start with medium sequences for better diversification\n # Generate diverse initial sequences with structured patterns and higher sums\n initial_sequences = []\n for _ in range(5):\n base = np.zeros(n)\n # Generate structured patterns with different characteristics\n if np.random.rand() < 0.3: # Cosine-like\n base = np.cos(np.linspace(0, np.pi, n)) * 0.2 + 0.1\n elif np.random.rand() < 0.3: # Uniform\n base = np.random.uniform(0.01, 0.05, size=n)\n elif np.random.rand() < 0.3: # Exponential\n base = np.random.exponential(scale=0.01, size=n) * 0.15\n elif np.random.rand() < 0.3: # Periodic\n base = np.sin(np.linspace(0, 2 * np.pi, n)) * 0.05 + 0.1\n else: # Random\n base = np.random.rand(n)\n sum_base = np.sum(base)\n scale_factor = 0.5 / sum_base if sum_base > 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[1:]:\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 random_walk_prob = 0.15 # Initial probability for random walks\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using dynamic constraint selection and adaptive line search.\"\"\"\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 conv = np.convolve(sequence, sequence)\n max_b_val = np.max(conv)\n # Dynamic tight position selection with higher threshold\n threshold = 0.9 * max_b_val # Increased threshold to include more constraints\n tight_positions = np.where(conv >= threshold)[0]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n # Solve LP with enhanced constraint selection\n g_fun = solve_convolution_lp(sequence, max_b_val, tight_positions)\n if g_fun is None:\n return None\n\n sum_g = np.sum(g_fun)\n if sum_g <= 0.0:\n return None\n\n # Use golden-section search on t to find optimal step\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 # Find optimal t between 0 and 1\n low = 0.0\n high = 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 \"\"\"Solves LP with enhanced constraint selection.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n # Variables\n g = cp.Variable(n, nonneg=True)\n # Objective: maximize sum(g_j), equivalent to minimizing -sum(g_j)\n objective = cp.Minimize(-cp.sum(g))\n\n # Constraints: ensure max(convolution) <= rhs\n constraints = []\n for k in tight_positions:\n # Coefficients for the constraint: sum_{j=0}^{n-1} f[i] * g[j] <= rhs\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\n # Solve the LP with a more efficient solver\n problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.SCS, verbose=False)\n if problem.status == cp.OPTIMAL:\n return g.value\n except:\n return None\n return None\n\n def perturb_sequence(seq):\n \"\"\"Perturbation with structured strategies for local minima escape.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n\n conv = np.convolve(seq, seq)\n # Select top 100 indices with highest convolution values for targeted changes\n top_indices = np.argsort(conv)[-100:]\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 # Occasionally perturb a random element\n if np.random.rand() < 0.2:\n random_idx = np.random.randint(0, len(seq))\n indices.append(random_idx)\n # Use adaptive scaling based on sequence properties\n scale = 0.15 * np.std(seq)\n for idx in indices:\n # Add random variation with varying scale\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 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 None:\n # Strategic expansion: add a new element to maintain pattern\n if len(current_sequence) < 1000:\n n = len(current_sequence)\n new_seq = current_sequence + [np.mean(current_sequence)]\n # Do not normalize to maintain increased sum\n current_sequence = new_seq\n print(f\"Expanded sequence to length {len(current_sequence)}\")\n else:\n current_sequence = perturb_sequence(current_sequence)\n else:\n current_sequence = h_function\n\n # Adjust random walk probability to explore less as solution improves\n if random_walk_prob > 0.05:\n random_walk_prob *= 0.995\n\n # Occasionally try random walk for exploration\n if np.random.rand() < random_walk_prob:\n current_sequence = random_walk(current_sequence)\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": 269.534353164956,
65 "env/all/time/policy/min": 94.53698515892029,
66 "env/all/time/policy/max": 352.4924826622009,
67 "env/all/time/env_step": 2843.082124957349,
68 "env/all/time/env_step/min": 0.009615898132324219,
69 "env/all/time/env_step/max": 5356.084791183472,
70 "env/all/time/reward_compute": 3.7671998143196106e-07,
71 "env/all/time/reward_compute/min": 2.0489096641540527e-07,
72 "env/all/time/reward_compute/max": 1.1771917343139648e-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": 0.01780363917350769,
77 "advantage/min": -1.0,
78 "advantage/max": 5.847777366638184,
79 "time/assemble_training_data": 5.305804967880249,
80 "time/kl_vs_base": 85.25529026985168,
81 "kl_policy_base": 0.0009426346514374018,
82 "time/train": 522.1252269744873,
83 "time/save_checkpoint": 23.07784938812256,
84 "time/total": 6350.327192783356
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