Views
No views yet
ac1. Checkpoint saved
after training step 35 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 35,
3 "progress/batch": 35,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.72,
6 "puct/buffer_size": 568,
7 "puct/sampled_size": 8,
8 "puct/T": 17920,
9 "puct/scale_last": 0.5749611129912189,
10 "puct/buffer_value/mean": -1.5242069357788508,
11 "puct/buffer_value/std": 0.08338301256275364,
12 "puct/buffer_value/min": -2.0797874683482562,
13 "puct/buffer_value/max": -1.5048263553570373,
14 "puct/buffer_timestep/mean": 16.746478873239436,
15 "puct/buffer_timestep/std": 10.249999395203218,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 34.0,
18 "puct/buffer_construction_len/mean": 1105.6161971830986,
19 "puct/buffer_construction_len/std": 692.03938045739,
20 "puct/buffer_construction_len/min": 1000.0,
21 "puct/buffer_construction_len/max": 7850.0,
22 "puct/sampled_value/mean": -1.504881400616859,
23 "puct/sampled_value/std": 3.660691210646036e-05,
24 "puct/sampled_value/min": -1.5049203859813658,
25 "puct/sampled_value/max": -1.5048263553570373,
26 "puct/sampled_timestep/mean": 34.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 34.0,
29 "puct/sampled_timestep/max": 34.0,
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": 6369.811160564423,
35 "env/all/ac_tokens_per_turn": 8097.76171875,
36 "env/all/ob_tokens_per_turn": 3562.5,
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": 4146054,
41 "env/all/total_ob_tokens": 1824000,
42 "env/all/time/sampling_mean": 508.6990666957572,
43 "env/all/time/sampling_max": 732.5126094818115,
44 "env/all/time/env_step_mean": 2957.00360747613,
45 "env/all/time/env_step_max": 5657.012530326843,
46 "env/all/reward/mean": 0.5588707395935314,
47 "env/all/reward/max": 0.664535202455724,
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": 0.5588707395935314,
53 "env/all/correctness": 0.876953125,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 4.205837342371922,
57 "env/all/raw_score/min": 1.5048111667512079,
58 "env/all/raw_score/max": 1023.9763789196033,
59 "env/all/initial_raw_score": -1.5048814006168592,
60 "env/all/initial_raw_score/min": -1.5049203859813658,
61 "env/all/initial_raw_score/max": -1.5048263553570373,
62 "env/all/msg": "RuntimeError: Program execution failed: NameError: name 'max_b_val' is not defined",
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 Uses a combination of structured perturbation, adaptive LP-based optimization, and enhanced exploration.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Initialize from known 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 # Generate Gaussian-like initial sequences\n for _ in range(40): # Fewer but more structured initial sequences\n n = np.random.randint(256, 1024)\n x = np.linspace(0, 1, n)\n base = np.exp(-((x - 0.5)**2) * 10) * 0.1 + 0.02\n sum_base = np.sum(base)\n scale_factor = 0.8 / (sum_base if sum_base > 0.0 else 0.5)\n seq = [max(0.0, x * scale_factor) for x in base]\n initial_sequences.append(seq)\n\n # Choose the best from initial sequences\n best_sequence = initial_sequences[0]\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\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 \"\"\"Perturb the most impactful positions in the convolution.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n conv = np.convolve(seq, seq)\n top_indices = np.argsort(conv)[-250:] # Expand more influential indices\n indices = np.random.choice(top_indices, size=min(70, len(top_indices)), replace=False)\n indices = [min(idx, len(seq)-1) for idx in indices]\n scale = 0.3 * np.std(seq) # Increased perturbation scale\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 \"\"\"Small random perturbations to all elements for exploration.\"\"\"\n scale = 0.2 * 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 \"\"\"Optimizes using dynamic LP and gradient-like blending.\"\"\"\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_scale = 0.8 # Reduced threshold scale to find tighter solutions\n threshold = max(0.5 * max_b_val, 0.05 * 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 g_fun = solve_convolution_lp(sequence, threshold, tight_positions, threshold_scale=threshold_scale)\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 # Golden-section search with increased 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): # More iterations for better optimization\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=1.0):\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 * max_b_val)\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 def expand_sequence(seq):\n \"\"\"Add elements carefully, balancing with current sequence.\"\"\"\n n = len(seq)\n if n >= 1500:\n return seq\n new_val = 0.9 * np.mean(seq)\n new_seq = seq + [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 return new_seq\n\n def structured_balance(seq):\n \"\"\"Adjust the top few peak positions in the convolution.\"\"\"\n n = len(seq)\n conv = np.convolve(seq, seq)\n peak_indices = np.argsort(conv)[-10:] # Focus on top contributing positions\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 return [max(0.0, min(1000.0, x)) for x in 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 not None:\n current_sequence = h_function\n else:\n # Expand sequence only if under limit\n if len(current_sequence) < 1500:\n current_sequence = expand_sequence(current_sequence)\n else:\n current_sequence = perturb_sequence(current_sequence)\n\n # Occasionally random walk for exploration\n if np.random.rand() < 0.02:\n current_sequence = random_walk(current_sequence)\n\n # Structured balance on top contributing positions\n if np.random.rand() < 0.02:\n current_sequence = structured_balance(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": 508.6990666957572,
65 "env/all/time/policy/min": 215.55411648750305,
66 "env/all/time/policy/max": 732.5126094818115,
67 "env/all/time/env_step": 2957.00360747613,
68 "env/all/time/env_step/min": 2.1556906700134277,
69 "env/all/time/env_step/max": 5657.012530326843,
70 "env/all/time/reward_compute": 7.14324414730072e-07,
71 "env/all/time/reward_compute/min": 2.123415470123291e-07,
72 "env/all/time/reward_compute/max": 2.1941959857940674e-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": 0.01900108903646469,
77 "advantage/min": -1.0,
78 "advantage/max": 7.273820877075195,
79 "time/assemble_training_data": 9.188367366790771,
80 "time/kl_vs_base": 141.58192539215088,
81 "kl_policy_base": 0.0010545337572693825,
82 "time/train": 1119.775083065033,
83 "time/save_checkpoint": 20.548264503479004,
84 "time/total": 7664.873247623444
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