Views
No views yet
erdos. Checkpoint saved
after training step 10 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.1{
2 "step": 10,
3 "progress/batch": 10,
4 "optim/lr": 4e-05,
5 "progress/done_frac": 0.22,
6 "puct/buffer_size": 168,
7 "puct/sampled_size": 8,
8 "puct/T": 5120,
9 "puct/scale_last": 0.030430701760354173,
10 "puct/buffer_value/mean": -0.3881391048001889,
11 "puct/buffer_value/std": 0.02540007298228171,
12 "puct/buffer_value/min": -0.5130522804051018,
13 "puct/buffer_value/max": -0.38094895199708134,
14 "puct/buffer_timestep/mean": 4.238095238095238,
15 "puct/buffer_timestep/std": 3.0379310300450633,
16 "puct/buffer_timestep/min": -1.0,
17 "puct/buffer_timestep/max": 9.0,
18 "puct/buffer_construction_len/mean": 66.17857142857143,
19 "puct/buffer_construction_len/std": 22.594595178770454,
20 "puct/buffer_construction_len/min": 40.0,
21 "puct/buffer_construction_len/max": 128.0,
22 "puct/sampled_value/mean": -0.38099582095832396,
23 "puct/sampled_value/std": 1.8981220310375933e-05,
24 "puct/sampled_value/min": -0.3810088231542323,
25 "puct/sampled_value/max": -0.38094895199708134,
26 "puct/sampled_timestep/mean": 9.0,
27 "puct/sampled_timestep/std": 0.0,
28 "puct/sampled_timestep/min": 9.0,
29 "puct/sampled_timestep/max": 9.0,
30 "puct/sampled_construction_len/mean": 77.0,
31 "puct/sampled_construction_len/std": 14.798648586948742,
32 "puct/sampled_construction_len/min": 64.0,
33 "puct/sampled_construction_len/max": 100.0,
34 "time/sampling": 1926.7340815067291,
35 "env/all/ac_tokens_per_turn": 8546.462890625,
36 "env/all/ob_tokens_per_turn": 1622.75,
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": 4375789,
41 "env/all/total_ob_tokens": 830848,
42 "env/all/time/sampling_mean": 466.33031401829794,
43 "env/all/time/sampling_max": 649.3448717594147,
44 "env/all/time/env_step_mean": 182.97795340325683,
45 "env/all/time/env_step_max": 1287.378674507141,
46 "env/all/reward/mean": 0.687151395441632,
47 "env/all/reward/max": 2.6250361915828053,
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": 0.687151395441632,
53 "env/all/correctness": 0.263671875,
54 "env/all/correctness/min": 0.0,
55 "env/all/correctness/max": 1.0,
56 "env/all/raw_score": 0.38407732491279734,
57 "env/all/raw_score/min": 0.3809471187315985,
58 "env/all/raw_score/max": 0.5108728990837365,
59 "env/all/initial_raw_score": -0.38099582095832396,
60 "env/all/initial_raw_score/min": -0.3810088231542323,
61 "env/all/initial_raw_score/max": -0.38094895199708134,
62 "env/all/msg": "Success; raw_score=0.3810273435542479",
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\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Generate a random initial guess that satisfies the integral constraint\n h_initial = np.random.uniform(0.0, 1.0, n_points)\n sum_h = np.sum(h_initial)\n h_initial[-1] = n_points / 2 - sum_h\n h_initial[-1] = np.clip(h_initial[-1], 0.0, 1.0)\n sum_h = np.sum(h_initial)\n if abs(sum_h - n_points / 2) > 1e-8:\n h_initial[0] = n_points / 2 - sum_h + h_initial[0]\n h_initial[0] = np.clip(h_initial[0], 0.0, 1.0)\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 max_corr = np.max(corr)\n return 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 # Perform basin-hopping with increased iterations\n result_bh = basinhopping(\n objective,\n h_initial,\n niter=1000,\n T=0.1,\n stepsize=0.1,\n minimizer_kwargs={\n 'method': 'SLSQP',\n 'bounds': bounds,\n 'constraints': [{'type': 'eq', 'fun': constraint_func}],\n 'options': {'ftol': 1e-10, 'maxiter': 2000, 'disp': False}\n },\n seed=seed\n )\n\n best_h = result_bh.x\n best_c5 = result_bh.fun\n\n # Refine using 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': 2000}\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):\n h_perturbed = best_h.copy()\n h_perturbed[0::5] += np.random.uniform(-0.02, 0.02, 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 smaller perturbations\n for _ in range(50):\n h_perturbed = best_h.copy()\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": 466.33031401829794,
65 "env/all/time/policy/min": 200.05661249160767,
66 "env/all/time/policy/max": 649.3448717594147,
67 "env/all/time/env_step": 182.97795340325683,
68 "env/all/time/env_step/min": 0.006216764450073242,
69 "env/all/time/env_step/max": 1287.378674507141,
70 "env/all/time/reward_compute": 3.1990930438041687e-07,
71 "env/all/time/reward_compute/min": 2.905726432800293e-07,
72 "env/all/time/reward_compute/max": 4.246830940246582e-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": 0.029055263847112656,
77 "advantage/min": -0.9647364616394043,
78 "advantage/max": 5.021705150604248,
79 "time/assemble_training_data": 8.075253248214722,
80 "time/kl_vs_base": 116.01116228103638,
81 "kl_policy_base": 0.0007771423552185297,
82 "time/train": 932.8687388896942,
83 "time/save_checkpoint": 9.767582178115845,
84 "time/total": 2996.171240091324
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