Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step35
ttt-discover-v3-qwen3-8b-erdos-s2-step35 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 34 (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 34 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 34,
"progress/batch": 34,
"optim/lr": 4e-05,
"progress/done_frac": 0.7,
"puct/buffer_size": 552,
"puct/sampled_size": 8,
"puct/T": 17408,
"puct/scale_last": 0.2790750662310147,
"puct/buffer_value/mean": -0.38409594371687866,
"puct/buffer_value/std": 0.01986572310282836,
"puct/buffer_value/min": -0.66,
"puct/buffer_value/max": -0.38092493376898534,
"puct/buffer_timestep/mean": 16.246376811594203,
"puct/buffer_timestep/std": 9.96138311376349,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 33.0,
"puct/buffer_construction_len/mean": 94.1231884057971,
"puct/buffer_construction_len/std": 33.1225114785347,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38092519042964373,
"puct/sampled_value/std": 1.262853386508055e-07,
"puct/sampled_value/min": -0.3809252643094563,
"puct/sampled_value/max": -0.38092493376898534,
"puct/sampled_timestep/mean": 33.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 33.0,
"puct/sampled_timestep/max": 33.0,
"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": 1823.6606664657593,
"env/all/ac_tokens_per_turn": 9229.177734375,
"env/all/ob_tokens_per_turn": 1790.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4725339,
"env/all/total_ob_tokens": 916544,
"env/all/time/sampling_mean": 526.547955406364,
"env/all/time/sampling_max": 755.3029522895813,
"env/all/time/env_step_mean": 43.98868204513565,
"env/all/time/env_step_max": 1100.1188917160034,
"env/all/reward/mean": 0.9968614979437365,
"env/all/reward/max": 2.625188567175501,
"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": 0.9968614979437365,
"env/all/correctness": 0.390625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3941227474914133,
"env/all/raw_score/min": 0.3809250071601663,
"env/all/raw_score/max": 0.5224411459633336,
"env/all/initial_raw_score": -0.3809251904296437,
"env/all/initial_raw_score/min": -0.3809252643094563,
"env/all/initial_raw_score/max": -0.38092493376898534,
"env/all/msg": "After normalization, h(x) is not in [0, 1]. Range: [2.4078937213121687e-05, 1.0000004368100628]",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\nfrom scipy.signal import correlate\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 100 # Optimal for speed and resolution balance\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n desired_sum = n_points / 2\n\n # Generate diverse initial population\n def generate_initial_pop():\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 initial_h[-1] += delta\n initial_h = np.clip(initial_h, 0.0, 1.0)\n return initial_h\n\n def random_initial_pop():\n h = np.random.uniform(0.0, 1.0, n_points)\n sum_h = np.sum(h)\n delta = desired_sum - sum_h\n h[-1] += delta\n h = np.clip(h, 0.0, 1.0)\n return h\n\n initial_population = [generate_initial_pop(), random_initial_pop()]\n initial_population.extend([random_initial_pop() for _ in range(2)]) # More diversity\n\n population = initial_population\n for _ in range(len(population), 200): # Ensure at least 200 individuals\n population.append(random_initial_pop())\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 generations = 1000\n mutation_decay_rate = (0.05 - 0.5) / generations\n mutation_rate = 0.5 # Starting rate\n\n for gen in range(generations):\n mutation_rate = 0.5 + mutation_decay_rate * gen\n\n # Evaluate and sort\n fitness = [evaluate_erdos(h) for h in population]\n best_idx = np.argmin(fitness)\n\n if fitness[best_idx] < best_c5:\n best_c5 = fitness[best_idx]\n best_h = population[best_idx].copy()\n\n # Select elites (20% of population)\n elite_count = int(0.2 * len(population))\n elite_indices = np.argsort(fitness)[:elite_count]\n elites = [population[i].copy() for i in elite_indices]\n\n # Create new population\n new_population = elites.copy()\n while len(new_population) < len(population):\n # Tournament selection\n idxs = np.random.choice(range(len(population)), size=2, replace=False)\n parent1, parent2 = population[idxs[0]], population[idxs[1]]\n child = (parent1 + parent2) / 2\n\n # Mutation\n if np.random.rand() < mutation_rate:\n num_mutations = np.random.randint(1, 11) # Mutation steps: 1 to 10\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 new_val = child[i] + delta\n new_val = np.clip(new_val, 0.0, 1.0)\n child[i] = new_val\n # Adjust sum to meet constraint\n sum_changes = np.sum(child) - np.sum(parent1)\n child[-1] = np.clip(child[-1] - sum_changes, 0.0, 1.0)\n\n new_population.append(child.copy())\n\n population = new_population\n\n # Final refinement with SLSQP\n if best_h is not None:\n h_array = np.array(best_h, dtype=np.float64)\n corr = correlate(h_array, 1 - h_array, mode='full')\n max_c5 = np.max(corr) * dx\n\n def objective(h):\n h_array = np.array(h, dtype=np.float64)\n corr = correlate(h_array, 1 - h_array, mode='full')\n return np.max(corr) * dx\n\n constraint_func = lambda h: np.sum(h) - desired_sum\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n res = minimize(\n fun=objective,\n x0=best_h,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={\n 'ftol': 1e-8,\n 'maxiter': 200,\n 'disp': False\n }\n )\n\n if res.success:\n new_c5 = evaluate_erdos(res.x)\n if new_c5 < best_c5:\n best_c5 = new_c5\n best_h = res.x\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 526.547955406364,
"env/all/time/policy/min": 193.25851845741272,
"env/all/time/policy/max": 755.3029522895813,
"env/all/time/env_step": 43.98868204513565,
"env/all/time/env_step/min": 0.007379055023193359,
"env/all/time/env_step/max": 1100.1188917160034,
"env/all/time/reward_compute": 4.0372833609580994e-07,
"env/all/time/reward_compute/min": 2.0489096641540527e-07,
"env/all/time/reward_compute/max": 9.164214134216309e-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": 0.02299928292632103,
"advantage/min": -1.0,
"advantage/max": 4.436617851257324,
"time/assemble_training_data": 10.438305616378784,
"time/kl_vs_base": 135.09484457969666,
"kl_policy_base": 0.0009138140012510121,
"time/train": 1030.1549615859985,
"time/save_checkpoint": 19.04296064376831,
"time/total": 3020.0868887901306
}
[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