Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step25
ttt-discover-v3-qwen3-8b-erdos-s2-step25 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 24 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 10, 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 24 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 24,
"progress/batch": 24,
"optim/lr": 4e-05,
"progress/done_frac": 0.5,
"puct/buffer_size": 392,
"puct/sampled_size": 8,
"puct/T": 12288,
"puct/scale_last": 0.1190747377095297,
"puct/buffer_value/mean": -0.38464560348641297,
"puct/buffer_value/std": 0.018902758794518353,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.3809252622904703,
"puct/buffer_timestep/mean": 11.244897959183673,
"puct/buffer_timestep/std": 7.075484009534579,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 23.0,
"puct/buffer_construction_len/mean": 84.3265306122449,
"puct/buffer_construction_len/std": 24.659218382354595,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38092531104281513,
"puct/sampled_value/std": 1.957450139746277e-08,
"puct/sampled_value/min": -0.38092532815402236,
"puct/sampled_value/max": -0.3809252622904703,
"puct/sampled_timestep/mean": 23.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 23.0,
"puct/sampled_timestep/max": 23.0,
"puct/sampled_construction_len/mean": 112.5,
"puct/sampled_construction_len/std": 33.071891388307385,
"puct/sampled_construction_len/min": 100.0,
"puct/sampled_construction_len/max": 200.0,
"time/sampling": 2406.5531299114227,
"env/all/ac_tokens_per_turn": 9245.125,
"env/all/ob_tokens_per_turn": 1512.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4733504,
"env/all/total_ob_tokens": 774208,
"env/all/time/sampling_mean": 508.78577444376424,
"env/all/time/sampling_max": 766.6783347129822,
"env/all/time/env_step_mean": 397.1112810089253,
"env/all/time/env_step_max": 1660.9369277954102,
"env/all/reward/mean": 1.1222312118433535,
"env/all/reward/max": 2.6251867873772294,
"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": 1.1222312118433535,
"env/all/correctness": 0.439453125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.39418579484806354,
"env/all/raw_score/min": 0.38092526541595606,
"env/all/raw_score/max": 0.675648725855192,
"env/all/initial_raw_score": -0.38092531104281513,
"env/all/initial_raw_score/min": -0.38092532815402236,
"env/all/initial_raw_score/max": -0.3809252622904703,
"env/all/msg": "Evaluation timed out after 1100 seconds.",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 200 # Increased number of samples for better resolution\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Generate multiple initial guesses\n initial_guesses = []\n\n # Include the initial guess from the provided variable\n if 'initial_h_values' in globals() and len(initial_h_values) == n_points:\n initial_guesses.append(np.array(initial_h_values, dtype=np.float64))\n\n # Additional randomized initial guesses to prevent local minima\n for _ in range(4):\n h = np.random.rand(n_points)\n h[-1] = (n_points / 2) - np.sum(h[:-1])\n h[-1] = np.clip(h[-1], 0.0, 1.0)\n initial_guesses.append(h)\n\n best_c5 = float('inf')\n best_h = None\n\n for init_h in initial_guesses:\n # Adjust for the sum constraint (ensures accuracy)\n current_sum = np.sum(init_h)\n desired_sum = n_points / 2\n delta = desired_sum - current_sum\n init_h[-1] += delta\n init_h[-1] = np.clip(init_h[-1], 0.0, 1.0)\n\n # Define objective function\n def objective(h):\n # Ensure the sum constraint is maintained\n adjusted_h = h.copy()\n sum_adj = np.sum(adjusted_h)\n delta = (n_points / 2) - sum_adj\n adjusted_h[-1] += delta\n adjusted_h[-1] = np.clip(adjusted_h[-1], 0.0, 1.0)\n\n # Compute the overlap using NumPy's correlate\n corr = np.correlate(adjusted_h, 1 - adjusted_h, mode=\"full\")\n max_overlap = np.max(corr) * dx\n return max_overlap\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Run Differential Evolution\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n maxiter=200,\n popsize=100,\n mutation=(0.6, 1.0),\n recombination=0.9,\n tol=1e-10,\n seed=seed,\n x0=init_h\n )\n\n # Refinement with SLSQP\n def constraint_func(h):\n return np.sum(h) - n_points / 2\n\n res = minimize(\n fun=objective,\n x0=result_de.x,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-8, 'maxiter': 300, 'disp': False}\n )\n\n current_c5 = res.fun\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_h = res.x\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 508.78577444376424,
"env/all/time/policy/min": 234.64161205291748,
"env/all/time/policy/max": 766.6783347129822,
"env/all/time/env_step": 397.1112810089253,
"env/all/time/env_step/min": 0.004807949066162109,
"env/all/time/env_step/max": 1660.9369277954102,
"env/all/time/reward_compute": 3.869645297527313e-07,
"env/all/time/reward_compute/min": 2.1979212760925293e-07,
"env/all/time/reward_compute/max": 7.37607479095459e-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": 0.02196202054619789,
"advantage/min": -1.0,
"advantage/max": 3.217851161956787,
"time/assemble_training_data": 8.15198040008545,
"time/kl_vs_base": 132.3160765171051,
"kl_policy_base": 0.000888441689312458,
"time/train": 1001.0588793754578,
"time/save_checkpoint": 18.47793936729431,
"time/total": 3569.383045196533
}
[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