Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step16
ttt-discover-v3-qwen3-8b-erdos-s2-step16 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 15 (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 15 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 15,
"progress/batch": 15,
"optim/lr": 4e-05,
"progress/done_frac": 0.32,
"puct/buffer_size": 248,
"puct/sampled_size": 8,
"puct/T": 7680,
"puct/scale_last": 0.11906820318179134,
"puct/buffer_value/mean": -0.38679137433072686,
"puct/buffer_value/std": 0.023499704761546017,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.38093179681820866,
"puct/buffer_timestep/mean": 6.741935483870968,
"puct/buffer_timestep/std": 4.479110956828331,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 14.0,
"puct/buffer_construction_len/mean": 74.8225806451613,
"puct/buffer_construction_len/std": 25.98512241434813,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38093789334964523,
"puct/sampled_value/std": 2.8131707928613054e-06,
"puct/sampled_value/min": -0.38094108128090803,
"puct/sampled_value/max": -0.38093179681820866,
"puct/sampled_timestep/mean": 14.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 14.0,
"puct/sampled_timestep/max": 14.0,
"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": 2659.43435049057,
"env/all/ac_tokens_per_turn": 9272.244140625,
"env/all/ob_tokens_per_turn": 1623.5,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4747389,
"env/all/total_ob_tokens": 831232,
"env/all/time/sampling_mean": 532.0063510988839,
"env/all/time/sampling_max": 776.1933326721191,
"env/all/time/env_step_mean": 557.4509519948624,
"env/all/time/env_step_max": 1894.4629139900208,
"env/all/reward/mean": 0.7592547820178215,
"env/all/reward/max": 2.6251651955314577,
"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": 0.7592547820178215,
"env/all/correctness": 0.296875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.39330315884662237,
"env/all/raw_score/min": 0.3809283985063274,
"env/all/raw_score/max": 0.5909020963368219,
"env/all/initial_raw_score": -0.38093789334964523,
"env/all/initial_raw_score/min": -0.38094108128090803,
"env/all/initial_raw_score/max": -0.38093179681820866,
"env/all/msg": "After normalization, h(x) is not in [0, 1]. Range: [7.042844907983182e-08, 1.0000000000000002]",
"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 # Define n_points for the problem\n n_points = 100\n dx = 2.0 / n_points\n\n # Use the provided initial_h_values if available, otherwise use a fallback\n if 'initial_h_values' in globals():\n h_initial = np.array(initial_h_values, dtype=np.float64)\n else:\n # Fallback initial guess: 5 intervals of 0.2 length\n h_initial = np.zeros(n_points, dtype=np.float64)\n for i in range(5):\n start = i * 0.4 # start of the interval\n end = start + 0.2 # end of the interval\n start_idx = int(start / dx)\n end_idx = int(end / dx)\n h_initial[start_idx:end_idx] = 1.0\n # Adjust the last element to meet the integral constraint\n h_initial[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_initial[:-1])))\n\n np.random.seed(seed)\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 # Initial basin-hopping with SLSQP\n result_bh = basinhopping(\n objective,\n h_initial,\n niter=500, # Sufficient iterations for exploration\n T=0.3, # Low temperature for precision\n stepsize=0.05,\n minimizer_kwargs={\n 'method': 'SLSQP',\n 'bounds': bounds,\n 'constraints': [{'type': 'eq', 'fun': constraint_func}],\n 'options': {'ftol': 1e-9, 'maxiter': 1000, 'disp': False}\n },\n seed=seed\n )\n\n best_h = result_bh.x\n best_c5 = result_bh.fun\n\n # Refinement loop with more aggressive perturbations and larger iterations\n for _ in range(1000): # Larger number for refinement\n h_perturbed = best_h.copy()\n h_array = np.array(h_perturbed)\n corr = np.correlate(h_array, 1 - h_array, mode='full')\n max_corr_idx = np.argmax(corr)\n k = max_corr_idx - n_points + 1\n\n # Determine which indices contribute to this k\n contributing_indices = set()\n for i in range(n_points):\n j = i + k\n if 0 <= j < n_points:\n contributing_indices.add(i)\n contributing_indices.add(j)\n\n # If there are contributing indices, perturb them\n contributing_indices = list(contributing_indices)\n if contributing_indices:\n perturb = np.random.uniform(-0.1, 0.1, size=len(contributing_indices))\n h_perturbed[contributing_indices] += perturb\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n\n # Maintain the integral constraint\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_perturbed[:-1])))\n\n # Local optimization to refine the solution\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-11, 'maxiter': 600}\n )\n\n if res_perturbed.fun < best_c5:\n best_h, best_c5 = res_perturbed.x, res_perturbed.fun\n\n # Return the result\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 532.0063510988839,
"env/all/time/policy/min": 149.4189350605011,
"env/all/time/policy/max": 776.1933326721191,
"env/all/time/env_step": 557.4509519948624,
"env/all/time/env_step/min": 0.0066487789154052734,
"env/all/time/env_step/max": 1894.4629139900208,
"env/all/time/reward_compute": 4.391185939311981e-07,
"env/all/time/reward_compute/min": 2.5331974029541016e-07,
"env/all/time/reward_compute/max": 1.2926757335662842e-06,
"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": 0.026607103645801544,
"advantage/min": -0.9586355090141296,
"advantage/max": 3.710892677307129,
"time/assemble_training_data": 9.695854425430298,
"time/kl_vs_base": 134.63509893417358,
"kl_policy_base": 0.0008100520935840905,
"time/train": 1016.7395775318146,
"time/save_checkpoint": 9.3303861618042,
"time/total": 3830.8729281425476
}
[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