Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step15
ttt-discover-v3-qwen3-8b-erdos-s2-step15 is a machine learning model from EdwardoSunny. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
LoRA adapter (rank 32, all-linear) from the upstream-faithful local reproduction of TTT-Discover on autoresearch-bench erdos. Checkpoint saved after training step 14 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 9, 2026
Repo size
349 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 100%
From the Hugging Face model README
LoRA adapter (rank 32, all-linear) from the upstream-faithful local
reproduction of TTT-Discover on autoresearch-bench erdos. Checkpoint saved
after training step 14 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 14,
"progress/batch": 14,
"optim/lr": 4e-05,
"progress/done_frac": 0.3,
"puct/buffer_size": 232,
"puct/sampled_size": 8,
"puct/T": 7168,
"puct/scale_last": 0.030441634417293917,
"puct/buffer_value/mean": -0.38616850544364983,
"puct/buffer_value/std": 0.021849101723020896,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.3809380193401416,
"puct/buffer_timestep/mean": 6.241379310344827,
"puct/buffer_timestep/std": 4.190754786874301,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 13.0,
"puct/buffer_construction_len/mean": 72.65517241379311,
"puct/buffer_construction_len/std": 24.669226763484186,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38094263478699125,
"puct/sampled_value/std": 2.653621496526063e-06,
"puct/sampled_value/min": -0.3809452345751634,
"puct/sampled_value/max": -0.3809380193401416,
"puct/sampled_timestep/mean": 13.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 13.0,
"puct/sampled_timestep/max": 13.0,
"puct/sampled_construction_len/mean": 100.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 100.0,
"puct/sampled_construction_len/max": 100.0,
"time/sampling": 2218.549244403839,
"env/all/ac_tokens_per_turn": 9578.58203125,
"env/all/ob_tokens_per_turn": 1558.75,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4904234,
"env/all/total_ob_tokens": 798080,
"env/all/time/sampling_mean": 542.8981395964511,
"env/all/time/sampling_max": 786.848443031311,
"env/all/time/env_step_mean": 296.3815986746922,
"env/all/time/env_step_max": 1452.8252725601196,
"env/all/reward/mean": 1.0531242225057882,
"env/all/reward/max": 2.6251417763002083,
"env/all/reward/min": 0.0,
"env/all/format": 1.0,
"env/all/format/min": 1.0,
"env/all/format/max": 1.0,
"env/all/reward": 1.0531242225057882,
"env/all/correctness": 0.4140625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3974088481791927,
"env/all/raw_score/min": 0.38093179681820866,
"env/all/raw_score/max": 1.0,
"env/all/initial_raw_score": -0.3809426347869912,
"env/all/initial_raw_score/min": -0.3809452345751634,
"env/all/initial_raw_score/max": -0.3809380193401416,
"env/all/msg": "C5 mismatch: reported 0.00000000, computed 0.50000000",
"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 = 200 # Increased for better approximation\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 h_initial = h_initial / np.sum(h_initial) * (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 # Basinhopping with better parameters\n result_bh = basinhopping(\n objective, h_initial,\n niter=500, T=0.5, stepsize=0.4,\n minimizer_kwargs={\n 'method': 'SLSQP',\n 'bounds': bounds,\n 'constraints': [{'type': 'eq', 'fun': constraint_func}],\n 'options': {'ftol': 1e-10, 'maxiter': 200}\n },\n seed=seed\n )\n\n best_h = result_bh.x\n best_c5 = result_bh.fun\n\n # Enhanced Refinement with adaptive strategies\n for _ in range(200): # More iterations for refinement\n h_perturbed = best_h.copy()\n\n # Identify most problematic shift\n corr = np.correlate(h_perturbed, 1 - h_perturbed, mode='full')\n max_corr_idx = np.argmax(corr)\n d = max_corr_idx\n\n # Compute gradient for this shift\n grad = np.zeros(n_points)\n for i in range(n_points):\n if i + d < n_points:\n grad[i] += (1 - h_perturbed[i + d])\n if i - d >= 0:\n grad[i] -= h_perturbed[i - d]\n\n # Apply a controlled perturbation\n step_size = 0.05 # Tuned step size\n h_perturbed -= step_size * grad\n\n # Clip and maintain constraint\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n sum_h = np.sum(h_perturbed)\n norm_factor = (n_points / 2) / sum_h\n h_perturbed *= norm_factor\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_perturbed[:-1])))\n\n # Local refinement using L-BFGS-B for targeted improvement\n res_perturbed = minimize(\n fun=objective,\n x0=h_perturbed,\n method='L-BFGS-B', # More efficient than SLSQP in this context\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 100}\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```",
"env/all/time/policy": 542.8981395964511,
"env/all/time/policy/min": 184.85670495033264,
"env/all/time/policy/max": 786.848443031311,
"env/all/time/env_step": 296.3815986746922,
"env/all/time/env_step/min": 0.005938291549682617,
"env/all/time/env_step/max": 1452.8252725601196,
"env/all/time/reward_compute": 4.246830940246582e-07,
"env/all/time/reward_compute/min": 2.7939677238464355e-07,
"env/all/time/reward_compute/max": 7.078051567077637e-07,
"env/all/by_group/frac_mixed": 1.0,
"env/all/by_group/frac_all_good": 0.0,
"env/all/by_group/frac_all_bad": 0.0,
"advantage/mean": 0.02088107541203499,
"advantage/min": -0.9999961853027344,
"advantage/max": 2.2518742084503174,
"time/assemble_training_data": 8.351229429244995,
"time/kl_vs_base": 136.32411217689514,
"kl_policy_base": 0.0007451729034073651,
"time/train": 1044.5903117656708,
"time/save_checkpoint": 11.290622234344482,
"time/total": 3420.349775791168
}
[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