Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step24
ttt-discover-v3-qwen3-8b-erdos-s2-step24 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 23 (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 23 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 23,
"progress/batch": 23,
"optim/lr": 4e-05,
"progress/done_frac": 0.48,
"puct/buffer_size": 376,
"puct/sampled_size": 8,
"puct/T": 11776,
"puct/scale_last": 0.11907469918514058,
"puct/buffer_value/mean": -0.3848039129550392,
"puct/buffer_value/std": 0.019284842161587323,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.3809253008148594,
"puct/buffer_timestep/mean": 10.74468085106383,
"puct/buffer_timestep/std": 6.786933895851247,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 22.0,
"puct/buffer_construction_len/mean": 83.3936170212766,
"puct/buffer_construction_len/std": 24.24244479152935,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.3809253333441207,
"puct/sampled_value/std": 1.7833119335569668e-08,
"puct/sampled_value/min": -0.38092534918379417,
"puct/sampled_value/max": -0.3809253014286148,
"puct/sampled_timestep/mean": 22.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 22.0,
"puct/sampled_timestep/max": 22.0,
"puct/sampled_construction_len/mean": 100.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 100.0,
"puct/sampled_construction_len/max": 100.0,
"time/sampling": 2126.7772920131683,
"env/all/ac_tokens_per_turn": 9553.337890625,
"env/all/ob_tokens_per_turn": 1422.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4891309,
"env/all/total_ob_tokens": 728064,
"env/all/time/sampling_mean": 533.0797228743322,
"env/all/time/sampling_max": 785.4683067798615,
"env/all/time/env_step_mean": 224.58262074179947,
"env/all/time/env_step_max": 1338.2503910064697,
"env/all/reward/mean": 1.179201197524773,
"env/all/reward/max": 2.625186808916845,
"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": 1.179201197524773,
"env/all/correctness": 0.46875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.4017267341613401,
"env/all/raw_score/min": 0.3809252622904703,
"env/all/raw_score/max": 1.0,
"env/all/initial_raw_score": -0.3809253333441208,
"env/all/initial_raw_score/min": -0.38092534918379417,
"env/all/initial_raw_score/max": -0.3809253014286148,
"env/all/msg": "Success; raw_score=0.3809253080338095",
"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 = 100 # Keep n_points small for speed\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Generate multiple initial guesses\n initial_guesses = []\n initial_guesses.append(initial_h_values)\n\n # Generate a random initial guess that satisfies the constraint\n h_random = np.random.rand(n_points)\n desired_sum = n_points / 2\n actual_sum = np.sum(h_random)\n h_random = h_random * (desired_sum / actual_sum)\n initial_guesses.append(h_random)\n\n best_c5 = float('inf')\n best_h = np.zeros(n_points)\n\n for init_h in initial_guesses:\n # Ensure constraint is met\n current_sum = np.sum(init_h)\n delta = desired_sum - current_sum\n if np.abs(delta) > 1e-6:\n init_h[-1] += delta\n init_h = np.clip(init_h, 0.0, 1.0)\n\n # Convert to numpy array\n init_h_array = np.array(init_h, dtype=np.float64)\n\n # Define objective with penalty for DE\n def objective_with_penalty(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n obj = np.max(corr) * dx\n penalty = 1e6 * (np.sum(h) - desired_sum) ** 2\n return obj + penalty\n\n # Bounds for DE\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Run Differential Evolution\n result_de = differential_evolution(\n objective_with_penalty,\n bounds,\n strategy='best1bin',\n popsize=30,\n mutation=(0.8, 1.2),\n recombination=0.9,\n tol=1e-6,\n maxiter=200,\n polish=True,\n disp=False\n )\n\n best_h_de = result_de.x\n best_c5_de = objective_with_penalty(best_h_de)\n\n # Local refinement with SLSQP\n def objective(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n return np.max(corr) * dx\n\n def constraint_func(h):\n return np.sum(h) - desired_sum\n\n result_slsqp = minimize(\n fun=objective,\n x0=best_h_de,\n method='SLSQP',\n bounds=bounds,\n constraints=[dict(type='eq', fun=constraint_func)],\n options=dict(ftol=1e-10, maxiter=2000)\n )\n\n current_c5 = result_slsqp.fun\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_h = result_slsqp.x\n\n # Fallback: minimize from initial guess\n result_initial = minimize(\n fun=objective,\n x0=init_h_array,\n method='SLSQP',\n bounds=bounds,\n constraints=[dict(type='eq', fun=constraint_func)],\n options=dict(ftol=1e-10, maxiter=2000)\n )\n\n current_c5_initial = result_initial.fun\n if current_c5_initial < best_c5:\n best_c5 = current_c5_initial\n best_h = result_initial.x\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 533.0797228743322,
"env/all/time/policy/min": 200.22980403900146,
"env/all/time/policy/max": 785.4683067798615,
"env/all/time/env_step": 224.58262074179947,
"env/all/time/env_step/min": 0.005672931671142578,
"env/all/time/env_step/max": 1338.2503910064697,
"env/all/time/reward_compute": 3.320164978504181e-07,
"env/all/time/reward_compute/min": 3.0547380447387695e-07,
"env/all/time/reward_compute/max": 3.762543201446533e-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": 0.02103484608232975,
"advantage/min": -1.0,
"advantage/max": 2.745429754257202,
"time/assemble_training_data": 8.441624164581299,
"time/kl_vs_base": 130.246031999588,
"kl_policy_base": 0.0009025484905578196,
"time/train": 1025.4901387691498,
"time/save_checkpoint": 14.584825038909912,
"time/total": 3306.970910549164
}
[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