Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step36
ttt-discover-v3-qwen3-8b-erdos-s2-step36 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 35 (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 35 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 35,
"progress/batch": 35,
"optim/lr": 4e-05,
"progress/done_frac": 0.72,
"puct/buffer_size": 568,
"puct/sampled_size": 8,
"puct/T": 17920,
"puct/scale_last": 0.2790750662310147,
"puct/buffer_value/mean": -0.3840084471634505,
"puct/buffer_value/std": 0.01959070866686422,
"puct/buffer_value/min": -0.66,
"puct/buffer_value/max": -0.38092493376898534,
"puct/buffer_timestep/mean": 16.746478873239436,
"puct/buffer_timestep/std": 10.249999395203218,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 34.0,
"puct/buffer_construction_len/mean": 94.28873239436619,
"puct/buffer_construction_len/std": 32.66713923374864,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.3809252154230453,
"puct/sampled_value/std": 8.623096736197675e-08,
"puct/sampled_value/min": -0.3809252634005885,
"puct/sampled_value/max": -0.3809250071601663,
"puct/sampled_timestep/mean": 34.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 34.0,
"puct/sampled_timestep/max": 34.0,
"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": 1850.2210557460785,
"env/all/ac_tokens_per_turn": 9635.90625,
"env/all/ob_tokens_per_turn": 1636.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4933584,
"env/all/total_ob_tokens": 837696,
"env/all/time/sampling_mean": 540.0089238029905,
"env/all/time/sampling_max": 880.9571042060852,
"env/all/time/env_step_mean": 50.58918967563659,
"env/all/time/env_step_max": 1100.1233944892883,
"env/all/reward/mean": 0.8598906719575643,
"env/all/reward/max": 2.6251952457875602,
"env/all/reward/min": 0.0,
"env/all/format": 0.998046875,
"env/all/format/min": 0.0,
"env/all/format/max": 1.0,
"env/all/reward": 0.8598906719575643,
"env/all/correctness": 0.33984375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.39968652792109277,
"env/all/raw_score/min": 0.3809240380701843,
"env/all/raw_score/max": 0.8,
"env/all/initial_raw_score": -0.38092521542304536,
"env/all/initial_raw_score/min": -0.3809252634005885,
"env/all/initial_raw_score/max": -0.3809250071601663,
"env/all/msg": "RuntimeError: Program execution failed: ValueError: operands could not be broadcast together with shapes (200,) (200,) (100,) ",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.signal import correlate\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 200 # Increased resolution for better approximation\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n desired_sum = n_points / 2\n population_size = 200 # Larger population for better exploration\n generations = 800 # Balanced number of generations\n\n # Initial population\n initial_h = np.array(initial_h_values, dtype=np.float64)\n initial_h = np.clip(initial_h, 0.0, 1.0)\n sum_h = np.sum(initial_h)\n delta = desired_sum - sum_h\n if abs(delta) > 1e-6:\n initial_h[-1] += delta\n initial_h = np.clip(initial_h, 0.0, 1.0)\n \n # Diverse population\n init_patterns = [\n initial_h.copy(),\n np.random.uniform(0, 1, n_points),\n np.random.permutation(n_points) / n_points, # Random permutation as pattern\n np.array([0 if i % 2 else 1 for i in range(n_points)], dtype=np.float64),\n np.sin(2 * np.pi * np.linspace(0, 1, n_points)) * 0.5 + 0.5,\n np.concatenate([initial_h[:n_points//2], initial_h[:n_points//2][::-1]]),\n ]\n\n # Finalize population\n population = [np.clip(p, 0.0, 1.0) for p in init_patterns]\n while len(population) < population_size:\n h = np.random.uniform(0.0, 1.0, n_points)\n h[-1] += desired_sum - np.sum(h)\n h = np.clip(h, 0.0, 1.0)\n population.append(h.copy())\n\n def evaluate_erdos(h):\n corr = correlate(h, 1 - h, mode='full')\n return np.max(corr) * dx\n\n best_c5 = float('inf')\n best_h = None\n\n for gen in range(generations):\n # Evaluate fitness\n fitness = [evaluate_erdos(h) for h in population]\n best_idx = np.argmin(fitness)\n if fitness[best_idx] < best_c5:\n best_c5 = fitness[best_idx]\n best_h = population[best_idx].copy()\n\n # Select elites\n elite_count = int(0.2 * population_size)\n elite_indices = np.argsort(fitness)[:elite_count]\n\n # Generate new population via crossover and mutation\n new_population = []\n\n # Elites\n for idx in elite_indices:\n new_population.append(population[idx].copy())\n\n # Generate offspring\n while len(new_population) < population_size:\n # Random parents\n idxs = np.random.choice(range(population_size), 2, replace=False)\n parent1, parent2 = population[idxs[0]], population[idxs[1]]\n \n # Uniform crossover\n child = np.where(np.random.rand(n_points) < 0.5, parent1, parent2)\n\n # Mutation\n if np.random.rand() < 0.2: # 20% mutation rate\n num_mutations = np.random.randint(1, 6) # 1-5 mutations\n indices = np.random.choice(range(n_points), num_mutations, replace=False)\n for i in indices:\n delta = np.random.uniform(-0.1, 0.1) # Larger mutation range\n child[i] += delta\n child[i] = np.clip(child[i], 0.0, 1.0)\n\n # Adjust sum constraint\n sum_changes = np.sum(child) - np.sum(parent1)\n child[-1] += sum_changes\n child = np.clip(child, 0.0, 1.0)\n\n new_population.append(child.copy())\n\n _ = population\n population = new_population\n\n # Final refinement\n if best_h is not None:\n best_h = best_h.copy()\n def objective(x):\n x = np.clip(x, 0.0, 1.0)\n x[-1] = np.clip(x[-1] + (desired_sum - np.sum(x)), 0.0, 1.0)\n return evaluate_erdos(x)\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n res = minimize(\n fun=objective,\n x0=best_h,\n method='L-BFGS-B',\n bounds=bounds,\n options={'ftol': 1e-8, 'maxiter': 200, 'disp': False}\n )\n\n if res.success:\n new_h = res.x\n new_c5 = evaluate_erdos(new_h)\n if new_c5 < best_c5:\n best_c5 = new_c5\n best_h = new_h\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 540.0089238029905,
"env/all/time/policy/min": 214.660710811615,
"env/all/time/policy/max": 880.9571042060852,
"env/all/time/env_step": 50.58918967563659,
"env/all/time/env_step/min": 0.006768941879272461,
"env/all/time/env_step/max": 1100.1233944892883,
"env/all/time/reward_compute": 3.6926940083503723e-07,
"env/all/time/reward_compute/min": 2.980232238769531e-07,
"env/all/time/reward_compute/max": 4.6193599700927734e-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": 0.02803300879895687,
"advantage/min": -0.9958076477050781,
"advantage/max": 8.400168418884277,
"time/assemble_training_data": 11.378679275512695,
"time/kl_vs_base": 143.84259867668152,
"kl_policy_base": 0.000868785020429641,
"time/train": 1059.0795938968658,
"time/save_checkpoint": 18.189764738082886,
"time/total": 3084.527283191681
}
[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