Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step43
ttt-discover-v3-qwen3-8b-erdos-s2-step43 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 42 (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 42 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 42,
"progress/batch": 42,
"optim/lr": 4e-05,
"progress/done_frac": 0.86,
"puct/buffer_size": 680,
"puct/sampled_size": 8,
"puct/T": 21504,
"puct/scale_last": 0.27911042634821415,
"puct/buffer_value/mean": -0.3841151170968237,
"puct/buffer_value/std": 0.019197200687244984,
"puct/buffer_value/min": -0.66,
"puct/buffer_value/max": -0.3808895736517859,
"puct/buffer_timestep/mean": 20.24705882352941,
"puct/buffer_timestep/std": 12.270390839312372,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 41.0,
"puct/buffer_construction_len/mean": 98.22941176470589,
"puct/buffer_construction_len/std": 35.68926786627108,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38089766059696967,
"puct/sampled_value/std": 4.662118756587482e-06,
"puct/sampled_value/min": -0.38090551677578943,
"puct/sampled_value/max": -0.3808895736517859,
"puct/sampled_timestep/mean": 41.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 41.0,
"puct/sampled_timestep/max": 41.0,
"puct/sampled_construction_len/mean": 200.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 200.0,
"puct/sampled_construction_len/max": 200.0,
"time/sampling": 2057.283857822418,
"env/all/ac_tokens_per_turn": 10061.59375,
"env/all/ob_tokens_per_turn": 1479.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 5151536,
"env/all/total_ob_tokens": 757312,
"env/all/time/sampling_mean": 585.1487099933438,
"env/all/time/sampling_max": 854.2471041679382,
"env/all/time/env_step_mean": 186.06084077991545,
"env/all/time/env_step_max": 1237.5054144859314,
"env/all/reward/mean": 0.6844641197842198,
"env/all/reward/max": 2.6254536291336596,
"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": 0.6844641197842198,
"env/all/correctness": 0.283203125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.4241979619800302,
"env/all/raw_score/min": 0.38088654952761103,
"env/all/raw_score/max": 0.98,
"env/all/initial_raw_score": -0.3808976605969697,
"env/all/initial_raw_score/min": -0.38090551677578943,
"env/all/initial_raw_score/max": -0.3808895736517859,
"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 norm, uniform\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n n_points = 200 # Optimal resolution under time constraint\n dx = 2.0 / n_points\n desired_sum = n_points / 2.0\n\n # Use the provided initial guess if available\n best_h = np.zeros(n_points)\n best_t = 0.5\n\n # Start with the provided initial_h_values if available\n if 'initial_h_values' in globals():\n initial_h = np.array(initial_h_values)\n # Ensure initial_h has the same length as n_points\n if initial_h.shape[0] > n_points:\n initial_h = initial_h[::2] # Downsample\n elif initial_h.shape[0] < n_points:\n initial_h = np.repeat(initial_h, n_points // len(initial_h))[:n_points]\n # Ensure sum is correct\n initial_h = np.clip(initial_h, 0, 1)\n initial_h[-1] += (desired_sum - np.sum(initial_h))\n initial_h = np.clip(initial_h, 0, 1)\n else:\n # Fallback: simple alternating pattern\n initial_h = np.zeros(n_points)\n for i in range(n_points):\n if i % 3 == 0:\n initial_h[i] = 1.0\n initial_h[-1] += (desired_sum - np.sum(initial_h))\n initial_h = np.clip(initial_h, 0, 1)\n\n # Generate diverse initial guesses\n num_perturbations = 30 # Reduce number of initial guesses to stay within budget\n\n def generate_perturbation(h):\n noise = norm.rvs(loc=0, scale=0.02, size=n_points)\n perturb = h + noise\n perturb = np.clip(perturb, 0, 1)\n sum_perturbed = np.sum(perturb)\n delta = desired_sum - sum_perturbed\n perturb[-1] += delta\n return perturb\n\n initial_guesses = []\n for _ in range(num_perturbations):\n initial_guesses.append(generate_perturbation(initial_h))\n\n # Optimize each initial guess\n for h in initial_guesses:\n def objective(x):\n return x[-1]\n\n # Constraint: \u2211h = desired_sum\n def constraint_sum(x):\n return np.sum(x[:-1]) - desired_sum\n\n # Constraint: t \u2265 overlap for selected worst shifts\n def constraints_func(x):\n h_vec = x[:-1]\n t = x[-1]\n cross_corr = np.correlate(h_vec, 1 - h_vec, mode='full')\n overlaps = cross_corr * dx\n # Select top 30 worst overlaps\n selected_overlaps = np.sort(overlaps)[-30:]\n return [t - overlap for overlap in selected_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, np.inf)]\n\n try:\n result = minimize(\n fun=objective,\n x0=np.concatenate((h, [0.5])),\n bounds=bounds,\n constraints=cons,\n method='SLSQP',\n options={\n 'ftol': 1e-8,\n 'maxiter': 200,\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\n\n return best_h, best_t, n_points\n```",
"env/all/time/policy": 585.1487099933438,
"env/all/time/policy/min": 229.9184684753418,
"env/all/time/policy/max": 854.2471041679382,
"env/all/time/env_step": 186.06084077991545,
"env/all/time/env_step/min": 0.0054247379302978516,
"env/all/time/env_step/max": 1237.5054144859314,
"env/all/time/reward_compute": 3.4598633646965027e-07,
"env/all/time/reward_compute/min": 2.682209014892578e-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": 0.029000893235206604,
"advantage/min": -0.870915412902832,
"advantage/max": 6.147963523864746,
"time/assemble_training_data": 11.593261480331421,
"time/kl_vs_base": 136.1246736049652,
"kl_policy_base": 0.0009036382543854415,
"time/train": 1089.5755598545074,
"time/save_checkpoint": 18.09086298942566,
"time/total": 3314.2613456249237
}
[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