Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step45
ttt-discover-v3-qwen3-8b-erdos-s2-step45 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 44 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 11, 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 44 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 44,
"progress/batch": 44,
"optim/lr": 4e-05,
"progress/done_frac": 0.9,
"puct/buffer_size": 712,
"puct/sampled_size": 8,
"puct/T": 22528,
"puct/scale_last": 0.2791134533744043,
"puct/buffer_value/mean": -0.3841729616309326,
"puct/buffer_value/std": 0.019267300365474638,
"puct/buffer_value/min": -0.66,
"puct/buffer_value/max": -0.38088654662559573,
"puct/buffer_timestep/mean": 21.247191011235955,
"puct/buffer_timestep/std": 12.847664850453647,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 43.0,
"puct/buffer_construction_len/mean": 102.9438202247191,
"puct/buffer_construction_len/std": 41.93517189169085,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 400.0,
"puct/sampled_value/mean": -0.38088725921262834,
"puct/sampled_value/std": 6.865592133955363e-07,
"puct/sampled_value/min": -0.38088865634629493,
"puct/sampled_value/max": -0.38088654662559573,
"puct/sampled_timestep/mean": 43.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 43.0,
"puct/sampled_timestep/max": 43.0,
"puct/sampled_construction_len/mean": 225.0,
"puct/sampled_construction_len/std": 66.14378277661477,
"puct/sampled_construction_len/min": 200.0,
"puct/sampled_construction_len/max": 400.0,
"time/sampling": 2178.227825164795,
"env/all/ac_tokens_per_turn": 10690.177734375,
"env/all/ob_tokens_per_turn": 1524.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 5473371,
"env/all/total_ob_tokens": 780352,
"env/all/time/sampling_mean": 641.389698874671,
"env/all/time/sampling_max": 963.806750535965,
"env/all/time/env_step_mean": 183.92318628681824,
"env/all/time/env_step_max": 1224.8555133342743,
"env/all/reward/mean": 0.5867250949079121,
"env/all/reward/max": 2.625465131656241,
"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": 0.5867250949079121,
"env/all/correctness": 0.240234375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.41975125123307294,
"env/all/raw_score/min": 0.3808848808113904,
"env/all/raw_score/max": 1.0,
"env/all/initial_raw_score": -0.38088725921262834,
"env/all/initial_raw_score/min": -0.38088865634629493,
"env/all/initial_raw_score/max": -0.38088654662559573,
"env/all/msg": "Computed C5 is not finite: nan",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\nfrom scipy.stats import uniform\nimport time\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n \n # Step 1: Choose higher resolution (n_points = 500)\n n_points = 500\n dx = 2.0 / n_points\n desired_sum = n_points / 2.0\n\n # Step 2: Use initial_h_values as starting point if available\n best_h = np.zeros(n_points)\n best_t = 0.5\n initial_h = None\n\n if 'initial_h_values' in globals():\n initial_h = np.array(initial_h_values)\n # Scale initial guess to n_points\n if initial_h.shape[0] < n_points:\n initial_h = np.repeat(initial_h, n_points // len(initial_h))[:n_points]\n # Adjust the last element to fix sum\n sum_h = np.sum(initial_h)\n delta = desired_sum - sum_h\n initial_h[-1] += delta\n initial_h = np.clip(initial_h, 0, 1)\n elif initial_h.shape[0] > n_points:\n initial_h = initial_h[::2] # take every other element\n # Adjust the last element\n sum_h = np.sum(initial_h)\n delta = desired_sum - sum_h\n initial_h[-1] += delta\n initial_h = np.clip(initial_h, 0, 1)\n # Ensure length matches n_points\n if len(initial_h) != n_points:\n initial_h = np.pad(initial_h, (0, n_points - len(initial_h)), 'constant')\n\n else:\n # Fallback to a structured pattern\n h = np.zeros(n_points)\n for i in range(n_points // 4):\n h[4*i] = 1.0\n sum_h = np.sum(h)\n delta = desired_sum - sum_h\n h[-1] += delta\n h = np.clip(h, 0, 1)\n initial_h = h\n\n # Step 3: Sample a subset of k for constraints to reduce computation\n num_constraints = 50 # Only use 50 k's instead of all 500\n k_values = np.random.uniform(-1.0, 1.0, size=num_constraints) # Sample shifts in [-1,1]\n k_values = np.sort(np.unique(k_values))\n k_values = np.concatenate([k_values, np.array([0.0])]) # Ensure k=0 is included\n\n def objective(x):\n return x[-1]\n\n def constraint_sum(x):\n h_vec = x[:-1]\n return np.sum(h_vec) - desired_sum\n\n def constraints_func(x):\n h_vec = x[:-1]\n t = x[-1]\n overlaps = []\n for k in k_values:\n # Avoid out-of-bound shifts due to periodicity (mod 2)\n k_mod = k % 2.0\n cross_corr = np.correlate(h_vec, 1 - h_vec, mode='full')\n # Compute the overlap for the current shift k_mod\n # Adjust the cross-correlation index to match k_mod\n # Use linear interpolation for shifts between points\n try:\n idx = int(k_mod / dx)\n if idx < 0:\n idx = 0\n elif idx >= len(cross_corr):\n idx = len(cross_corr) - 1\n overlap_idx = idx\n overlap = cross_corr[overlap_idx] * dx\n overlaps.append(overlap)\n except:\n overlaps.append(0.0)\n return [t - overlap for overlap in overlaps]\n\n cons = [\n {'type': 'eq', 'fun': constraint_sum},\n {'type': 'ineq', 'fun': constraints_func}\n ]\n\n bounds = [(0.0, 1.0) for _ in range(n_points)] + [(0.0, 1.0)]\n\n # Step 4: Run optimization with a single initial guess\n try:\n result = minimize(\n fun=objective,\n x0=np.concatenate((initial_h, [0.5])),\n bounds=bounds,\n constraints=cons,\n method='COBYLA',\n options={\n 'ftol': 1e-10,\n 'maxiter': 500,\n 'disp': False\n }\n )\n if result.success and result.x[-1] < best_t:\n best_t = result.x[-1]\n best_h = result.x[:-1]\n except:\n pass # Skip on failure\n\n return best_h, best_t, n_points\n```",
"env/all/time/policy": 641.389698874671,
"env/all/time/policy/min": 281.96221256256104,
"env/all/time/policy/max": 963.806750535965,
"env/all/time/env_step": 183.92318628681824,
"env/all/time/env_step/min": 0.009228944778442383,
"env/all/time/env_step/max": 1224.8555133342743,
"env/all/time/reward_compute": 4.2654573917388916e-07,
"env/all/time/reward_compute/min": 1.9744038581848145e-07,
"env/all/time/reward_compute/max": 1.2293457984924316e-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": 0.03199121356010437,
"advantage/min": -0.8827623128890991,
"advantage/max": 5.485356330871582,
"time/assemble_training_data": 8.947662115097046,
"time/kl_vs_base": 143.93750739097595,
"kl_policy_base": 0.0008192691602744162,
"time/train": 1165.6735730171204,
"time/save_checkpoint": 18.171627283096313,
"time/total": 3516.2749469280243
}
[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