Views
No views yet
erdos. Checkpoint saved
after training step 11 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 11,
3 "progress/batch": 11,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.24,
6 "puct/buffer_size": 184,
7 "puct/sampled_size": 8,
8 "puct/T": 5632,
9 "puct/scale_last": 0.030432535025837038,
10 "puct/buffer_value/mean": -0.3875207457605026,
11 "puct/buffer_value/std": 0.024353195854402904,
12 "puct/buffer_value/min": -0.5130522804051018,
13 "puct/buffer_value/max": -0.3809471187315985,
14 "puct/buffer_timestep/mean": 4.739130434782608,
15 "puct/buffer_timestep/std": 3.326015912853092,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 10.0,
18 "puct/buffer_construction_len/mean": 68.01086956521739,
19 "puct/buffer_construction_len/std": 24.57331292153153,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 200.0,
22 "puct/sampled_value/mean": -0.3809826776415831,
23 "puct/sampled_value/std": 2.1278994887352916e-05,
24 "puct/sampled_value/min": -0.38100206801482206,
25 "puct/sampled_value/max": -0.3809471187315985,
26 "puct/sampled_timestep/mean": 10.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 10.0,
29 "puct/sampled_timestep/max": 10.0,
30 "puct/sampled_construction_len/mean": 83.5,
31 "puct/sampled_construction_len/std": 14.203872711341791,
32 "puct/sampled_construction_len/min": 64.0,
33 "puct/sampled_construction_len/max": 100.0,
34 "time/sampling": 1962.5900192260742,
35 "env/all/ac_tokens_per_turn": 8653.662109375,
36 "env/all/ob_tokens_per_turn": 1644.125,
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": 4430675,
41 "env/all/total_ob_tokens": 841792,
42 "env/all/time/sampling_mean": 466.9482832849026,
43 "env/all/time/sampling_max": 687.9149513244629,
44 "env/all/time/env_step_mean": 197.57177604502067,
45 "env/all/time/env_step_max": 1296.0188472270966,
46 "env/all/reward/mean": 0.5930686875743132,
47 "env/all/reward/max": 2.6250673664885533,
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": 0.5930686875743132,
53 "env/all/correctness": 0.228515625,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.3863694555233757,
57 "env/all/raw_score/min": 0.38094259466071756,
58 "env/all/raw_score/max": 0.5257025002770781,
59 "env/all/initial_raw_score": -0.3809826776415831,
60 "env/all/initial_raw_score/min": -0.38100206801482206,
61 "env/all/initial_raw_score/max": -0.3809471187315985,
62 "env/all/msg": "Success; raw_score=0.38094671042947825",
63 "env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import basinhopping, minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 100 # Optimal for balance between accuracy and speed\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Generate a symmetric binary initial guess\n h_initial = np.zeros(n_points)\n i_start = int(0.5 / dx) # corresponds to x=0.5\n i_end = int(1.5 / dx) # corresponds to x=1.5\n h_initial[i_start:i_end+1] = 1.0\n h_initial[i_end] = 0.0 # Ensure sum is 50 (to satisfy sum(h) = n_points / 2)\n\n def objective(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n return np.max(corr) * dx\n\n def constraint_func(h):\n return np.sum(h) - n_points / 2\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Basin-hopping with enhanced parameters for better exploration\n result_bh = basinhopping(\n objective,\n h_initial,\n niter=700, # More basin-hopping\n T=0.15, # Slightly higher temperature for broader exploration\n stepsize=0.15, # Moderate stepsize for balance\n minimizer_kwargs={\n 'method': 'SLSQP',\n 'bounds': bounds,\n 'constraints': [{'type': 'eq', 'fun': constraint_func}],\n 'options': {'ftol': 1e-10, 'maxiter': 1500, 'disp': False}\n },\n seed=seed\n )\n\n best_h = result_bh.x\n best_c5 = result_bh.fun\n\n # Refine with SLSQP\n res_slsqp = minimize(\n fun=objective,\n x0=best_h,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 1500}\n )\n\n best_h, best_c5 = res_slsqp.x, res_slsqp.fun\n\n # Guided perturbations for deeper refinement\n for _ in range(200): # More extensive refinement\n h_perturbed = best_h.copy()\n # Small structured perturbation\n h_perturbed[0::5] += np.random.uniform(-0.015, 0.015, size=len(h_perturbed[0::5]))\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_perturbed[:-1])))\n\n res_perturbed = minimize(\n fun=objective,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 300}\n )\n\n if res_perturbed.fun < best_c5:\n best_h, best_c5 = res_perturbed.x, res_perturbed.fun\n\n # Final refinement with localized perturbations\n for _ in range(50):\n h_perturbed = best_h.copy()\n # Localized perturbation\n h_perturbed[20:-10] += np.random.uniform(-0.01, 0.01, size=len(h_perturbed[20:-10]))\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_perturbed[:-1])))\n\n res_perturbed = minimize(\n fun=objective,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 150}\n )\n\n if res_perturbed.fun < best_c5:\n best_h, best_c5 = res_perturbed.x, res_perturbed.fun\n\n return (best_h, best_c5, n_points)\n```",
64 "env/all/time/policy": 466.9482832849026,
65 "env/all/time/policy/min": 190.25409317016602,
66 "env/all/time/policy/max": 687.9149513244629,
67 "env/all/time/env_step": 197.57177604502067,
68 "env/all/time/env_step/min": 0.005814790725708008,
69 "env/all/time/env_step/max": 1296.0188472270966,
70 "env/all/time/reward_compute": 3.67872416973114e-07,
71 "env/all/time/reward_compute/min": 1.9371509552001953e-07,
72 "env/all/time/reward_compute/max": 7.189810276031494e-07,
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": 0.03056248277425766,
77 "advantage/min": -0.8696556091308594,
78 "advantage/max": 5.102476596832275,
79 "time/assemble_training_data": 9.039973020553589,
80 "time/kl_vs_base": 123.88618302345276,
81 "kl_policy_base": 0.0007670099730603397,
82 "time/train": 945.6399366855621,
83 "time/save_checkpoint": 15.091049432754517,
84 "time/total": 3057.400750875473
85}[2026-07-09T06:24:18+00:00] job=1812626 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:02:36+00:00] job=1813125 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:32:13+00:00] job=1813126 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T10:04:31+00:00] job=1813623 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=yes