Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step11
ttt-discover-v3-qwen3-8b-erdos-step11 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 10 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 9, 2026
Repo size
354 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 99%
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 10 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 10,
"progress/batch": 10,
"optim/lr": 4e-05,
"progress/done_frac": 0.22,
"puct/buffer_size": 168,
"puct/sampled_size": 8,
"puct/T": 5120,
"puct/scale_last": 0.06766262263675182,
"puct/buffer_value/mean": -0.38849651748867503,
"puct/buffer_value/std": 0.02721991562883291,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.380941803845533,
"puct/buffer_timestep/mean": 4.238095238095238,
"puct/buffer_timestep/std": 3.0379310300450633,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 9.0,
"puct/buffer_construction_len/mean": 77.30357142857143,
"puct/buffer_construction_len/std": 16.97012744030333,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38094790515625154,
"puct/sampled_value/std": 4.3553334598275435e-06,
"puct/sampled_value/min": -0.38095133431987294,
"puct/sampled_value/max": -0.380941803845533,
"puct/sampled_timestep/mean": 9.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 9.0,
"puct/sampled_timestep/max": 9.0,
"puct/sampled_construction_len/mean": 80.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 80.0,
"puct/sampled_construction_len/max": 80.0,
"time/sampling": 1815.2153310775757,
"env/all/ac_tokens_per_turn": 9334.28515625,
"env/all/ob_tokens_per_turn": 1432.5,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4779154,
"env/all/total_ob_tokens": 733440,
"env/all/time/sampling_mean": 509.1595449456945,
"env/all/time/sampling_max": 726.8818044662476,
"env/all/time/env_step_mean": 72.8695426704362,
"env/all/time/env_step_max": 1100.1204903125763,
"env/all/reward/mean": 1.2568637064606918,
"env/all/reward/max": 2.625073083533121,
"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": 1.2568637064606918,
"env/all/correctness": 0.490234375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3921426231153824,
"env/all/raw_score/min": 0.3809417650206355,
"env/all/raw_score/max": 0.5602527492679265,
"env/all/initial_raw_score": -0.38094790515625154,
"env/all/initial_raw_score/min": -0.38095133431987294,
"env/all/initial_raw_score/max": -0.380941803845533,
"env/all/msg": "Success; raw_score=0.38094180278183704",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import basinhopping, Bounds\nfrom scipy.signal import correlate\nimport random\n\ndef generate_initial_guesses(n_points, required_sum):\n guesses = []\n # Initial construction from provided values\n h_initial = initial_h_values.copy()\n h_initial = h_initial * required_sum / np.sum(h_initial)\n guesses.append(h_initial)\n \n # Random perturbation\n h_perturb = h_initial + np.random.normal(0, 0.05, n_points)\n h_perturb = np.clip(h_perturb, 0.0, 1.0)\n h_perturb = h_perturb * required_sum / np.sum(h_perturb)\n guesses.append(h_perturb)\n \n # Evenly spaced 1s\n h_even = np.zeros(n_points)\n for i in range(n_points):\n if i % 2 == 0:\n h_even[i] = 1.0\n h_even = h_even * required_sum / np.sum(h_even)\n guesses.append(h_even)\n \n # Cluster-based guess\n cluster_start = np.random.choice(n_points, size=3, replace=False)\n cluster_end = np.random.choice(n_points, size=3, replace=False)\n h_clusters = np.zeros(n_points)\n for i, j in zip(cluster_start, cluster_end):\n h_clusters[i:j] = 1.0\n h_clusters = h_clusters * required_sum / np.sum(h_clusters)\n guesses.append(h_clusters)\n \n # Additional random guesses\n for _ in range(4):\n h_random = np.random.rand(n_points)\n h_random = h_random * required_sum / np.sum(h_random)\n guesses.append(h_random)\n \n return guesses\n\ndef evaluate_C5(h, dx):\n h1 = 1.0 - h\n corr = correlate(h, h1, mode='full')\n max_corr = np.max(corr)\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n random.seed(seed)\n \n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = int(n_points / 2)\n \n initial_guesses = generate_initial_guesses(n_points, required_sum)\n \n best_h = None\n best_c5 = float('inf')\n \n for guess in initial_guesses:\n # Objective function with constraint penalty\n def objective(x):\n x = np.clip(x, 0.0, 1.0)\n c5_val = evaluate_C5(x, dx)\n penalty = 1000.0 * (np.sum(x) - required_sum)**2\n return c5_val + penalty\n \n bounds = Bounds(0.0, 1.0)\n \n # Perform Basin Hopping optimization\n result = basinhopping(\n objective,\n guess,\n niter=50, # Number of basin-hopping iterations\n T=1.0,\n stepsize=0.1,\n minimizer_kwargs={\n 'method': 'L-BFGS-B',\n 'bounds': bounds,\n 'options': {'ftol': 1e-8, 'maxiter': 100}\n }\n )\n \n current_h = result.x\n current_c5 = evaluate_C5(current_h, dx)\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_h = current_h\n \n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 509.1595449456945,
"env/all/time/policy/min": 236.45835638046265,
"env/all/time/policy/max": 726.8818044662476,
"env/all/time/env_step": 72.8695426704362,
"env/all/time/env_step/min": 0.004758358001708984,
"env/all/time/env_step/max": 1100.1204903125763,
"env/all/time/reward_compute": 3.413297235965729e-07,
"env/all/time/reward_compute/min": 1.862645149230957e-07,
"env/all/time/reward_compute/max": 6.109476089477539e-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": 0.019667334854602814,
"advantage/min": -1.0,
"advantage/max": 4.55954647064209,
"time/assemble_training_data": 6.48383355140686,
"time/kl_vs_base": 125.81721949577332,
"kl_policy_base": 0.0007744097383692861,
"time/train": 1005.7674596309662,
"time/save_checkpoint": 10.715528964996338,
"time/total": 2967.063055038452
}
[2026-07-09T06:24:18+00:00] job=1812624 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:31:44+00:00] job=1812955 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T07:59:00+00:00] job=1813123 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:28:32+00:00] job=1813124 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T09:35:59+00:00] job=1813609 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:45:57+00:00] job=1813622 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes