Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step13
ttt-discover-v3-qwen3-8b-erdos-step13 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 12 (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 12 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 12,
"progress/batch": 12,
"optim/lr": 4e-05,
"progress/done_frac": 0.26,
"puct/buffer_size": 200,
"puct/sampled_size": 8,
"puct/T": 6144,
"puct/scale_last": 0.06766266166080398,
"puct/buffer_value/mean": -0.38728845252743027,
"puct/buffer_value/std": 0.02510055698602651,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.38094176482148084,
"puct/buffer_timestep/mean": 5.24,
"puct/buffer_timestep/std": 3.6141942393844855,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 11.0,
"puct/buffer_construction_len/mean": 77.735,
"puct/buffer_construction_len/std": 15.584760986296837,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38094199423078323,
"puct/sampled_value/std": 2.290733134707792e-07,
"puct/sampled_value/min": -0.3809422831305527,
"puct/sampled_value/max": -0.38094176482148084,
"puct/sampled_timestep/mean": 11.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 11.0,
"puct/sampled_timestep/max": 11.0,
"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": 1524.4238317012787,
"env/all/ac_tokens_per_turn": 9003.404296875,
"env/all/ob_tokens_per_turn": 1716.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4609743,
"env/all/total_ob_tokens": 878656,
"env/all/time/sampling_mean": 497.62415861245245,
"env/all/time/sampling_max": 717.3696942329407,
"env/all/time/env_step_mean": 25.172586089000106,
"env/all/time/env_step_max": 1100.1439545154572,
"env/all/reward/mean": 1.3820075946893353,
"env/all/reward/max": 2.6250736542263016,
"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": 1.3820075946893353,
"env/all/correctness": 0.533203125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3881759139814203,
"env/all/raw_score/min": 0.38094168220357894,
"env/all/raw_score/max": 1.0,
"env/all/initial_raw_score": -0.38094199423078323,
"env/all/initial_raw_score/min": -0.3809422831305527,
"env/all/initial_raw_score/max": -0.38094176482148084,
"env/all/msg": "Success; raw_score=0.380941764785875",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize, Bounds\nfrom scipy.signal import correlate\n\ndef generate_initial_guesses(n_points, required_sum):\n guesses = []\n\n # Original initial guess\n h = initial_h_values.copy()\n h = h * required_sum / np.sum(h)\n guesses.append(h)\n\n # Perturbed random version\n perturbed = h + np.random.normal(0, 0.05, n_points)\n perturbed = np.clip(perturbed, 0.0, 1.0)\n perturbed = perturbed * required_sum / np.sum(perturbed)\n guesses.append(perturbed)\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\n h_even = h_even * required_sum / np.sum(h_even)\n guesses.append(h_even)\n\n # Random cluster-based guess\n cluster_start = np.random.choice(n_points, size=4, replace=False)\n cluster_end = np.random.choice(n_points, size=4, replace=False)\n h_clusters = np.zeros(n_points)\n for i, j in zip(sorted(cluster_start), sorted(cluster_end)):\n h_clusters[i:j] = 1\n h_clusters = h_clusters * required_sum / np.sum(h_clusters)\n guesses.append(h_clusters)\n\n # Random binary guess (exactly required_sum 1s)\n indices = np.random.choice(n_points, size=required_sum, replace=False)\n h_binary = np.zeros(n_points)\n h_binary[indices] = 1\n guesses.append(h_binary)\n\n # Randomly perturbed version of h_binary\n h_perturbed_binary = h_binary.copy()\n for i in range(n_points):\n if np.random.rand() < 0.1:\n h_perturbed_binary[i] = np.clip(h_perturbed_binary[i] + np.random.normal(0, 0.1), 0.0, 1.0)\n h_perturbed_binary = h_perturbed_binary * required_sum / np.sum(h_perturbed_binary)\n guesses.append(h_perturbed_binary)\n\n # Random initial guess\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 # Random even spacing\n h_even_rand = np.zeros(n_points)\n for i in range(n_points):\n if i % 4 == 0:\n h_even_rand[i] = 1\n h_even_rand = h_even_rand * required_sum / np.sum(h_even_rand)\n guesses.append(h_even_rand)\n\n # Random sparse binary\n indices_sparse = np.random.choice(n_points, size=required_sum, replace=False)\n h_sparse = np.zeros(n_points)\n h_sparse[indices_sparse] = 1\n guesses.append(h_sparse)\n\n # Perturbed version of sparse binary\n h_perturbed_sparse = h_sparse.copy()\n for i in range(n_points):\n if np.random.rand() < 0.1:\n h_perturbed_sparse[i] = np.clip(h_perturbed_sparse[i] + np.random.normal(0, 0.1), 0.0, 1.0)\n h_perturbed_sparse = h_perturbed_sparse * required_sum / np.sum(h_perturbed_sparse)\n guesses.append(h_perturbed_sparse)\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 n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = int(n_points / 2)\n\n def objective(x):\n h1 = 1.0 - x\n corr = correlate(x, h1, mode='full')\n max_corr = np.max(corr)\n constraint_violation = np.abs(np.sum(x) - required_sum)\n return max_corr * dx + 1e5 * constraint_violation ** 2\n\n initial_guesses = generate_initial_guesses(n_points, required_sum)\n best_h = None\n best_c5 = float('inf')\n\n for guess in initial_guesses:\n result = minimize(\n fun=objective,\n x0=guess,\n method='L-BFGS-B',\n bounds=Bounds(0.0, 1.0),\n options={'ftol': 1e-10, 'maxiter': 5000}\n )\n if result.success:\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": 497.62415861245245,
"env/all/time/policy/min": 237.45468044281006,
"env/all/time/policy/max": 717.3696942329407,
"env/all/time/env_step": 25.172586089000106,
"env/all/time/env_step/min": 0.006636142730712891,
"env/all/time/env_step/max": 1100.1439545154572,
"env/all/time/reward_compute": 3.520399332046509e-07,
"env/all/time/reward_compute/min": 2.0489096641540527e-07,
"env/all/time/reward_compute/max": 7.711350917816162e-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": 0.015620511025190353,
"advantage/min": -1.0,
"advantage/max": 2.564751625061035,
"time/assemble_training_data": 8.19255781173706,
"time/kl_vs_base": 127.94262766838074,
"kl_policy_base": 0.000779210589826107,
"time/train": 999.1997621059418,
"time/save_checkpoint": 3.1988413333892822,
"time/total": 2664.0516147613525
}
[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