Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step33
ttt-discover-v3-qwen3-8b-erdos-s2-step33 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 32 (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 32 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 32,
"progress/batch": 32,
"optim/lr": 4e-05,
"progress/done_frac": 0.66,
"puct/buffer_size": 520,
"puct/sampled_size": 8,
"puct/T": 16384,
"puct/scale_last": 0.27907477309284534,
"puct/buffer_value/mean": -0.3842834361045003,
"puct/buffer_value/std": 0.020452761360597583,
"puct/buffer_value/min": -0.66,
"puct/buffer_value/max": -0.3809252269071547,
"puct/buffer_timestep/mean": 15.246153846153845,
"puct/buffer_timestep/std": 9.384161401377355,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 31.0,
"puct/buffer_construction_len/mean": 93.76153846153846,
"puct/buffer_construction_len/std": 34.09337942636458,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38092525894115536,
"puct/sampled_value/std": 1.3650444243490477e-08,
"puct/sampled_value/min": -0.38092526668209503,
"puct/sampled_value/max": -0.3809252269071547,
"puct/sampled_timestep/mean": 31.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 31.0,
"puct/sampled_timestep/max": 31.0,
"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": 1831.7962024211884,
"env/all/ac_tokens_per_turn": 9463.78515625,
"env/all/ob_tokens_per_turn": 1727.75,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4845458,
"env/all/total_ob_tokens": 884608,
"env/all/time/sampling_mean": 530.2478166646324,
"env/all/time/sampling_max": 791.3377456665039,
"env/all/time/env_step_mean": 35.246664189267904,
"env/all/time/env_step_max": 1100.1197152137756,
"env/all/reward/mean": 0.999193683099558,
"env/all/reward/max": 2.6251872337375532,
"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": 0.999193683099558,
"env/all/correctness": 0.39453125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3975185675790633,
"env/all/raw_score/min": 0.3809252006472694,
"env/all/raw_score/max": 0.5257025002811964,
"env/all/initial_raw_score": -0.38092525894115536,
"env/all/initial_raw_score/min": -0.38092526668209503,
"env/all/initial_raw_score/max": -0.3809252269071547,
"env/all/msg": "Success; raw_score=0.3809252006472694",
"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 # Hyperparameters\n population_size = 100\n generations = 500 # Reduced from 1000 for faster execution\n elite_ratio = 0.2\n mutation_rate_start = 0.15\n mutation_rate_end = 0.01\n mutation_decay = (mutation_rate_end - mutation_rate_start) / generations\n mutation_steps = 3 # Number of mutations per individual in this version\n\n # Initialize population with diverse entries\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 population = [initial_h.copy()]\n for _ in range(population_size - 1):\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 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 mutation_rate = mutation_rate_start + mutation_decay * 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\n elite_count = int(elite_ratio * population_size)\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) < population_size:\n # Tournament selection\n idxs = np.random.choice(range(population_size), size=2, replace=False)\n parent1, parent2 = population[idxs[0]], population[idxs[1]]\n child = (parent1 + parent2) / 2\n\n # Mutation: Allow multiple elements to be mutated\n if np.random.rand() < mutation_rate:\n num_mutations = np.random.randint(1, 4) # 1 to 3 elements to mutate\n indices = np.random.choice(range(n_points), num_mutations, replace=False)\n # Apply mutations\n for i in indices:\n delta = np.random.uniform(-0.03, 0.03)\n new_val = child[i] + delta\n new_val = np.clip(new_val, 0.0, 1.0)\n child[i] = new_val\n # Adjust last element to maintain sum\n sum_changes = np.sum(child) - np.sum(parent1) # Using parent1's original sum\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-7,\n 'maxiter': 500,\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": 530.2478166646324,
"env/all/time/policy/min": 252.7631800174713,
"env/all/time/policy/max": 791.3377456665039,
"env/all/time/env_step": 35.246664189267904,
"env/all/time/env_step/min": 0.0073664188385009766,
"env/all/time/env_step/max": 1100.1197152137756,
"env/all/time/reward_compute": 2.915039658546448e-07,
"env/all/time/reward_compute/min": 1.564621925354004e-07,
"env/all/time/reward_compute/max": 5.178153514862061e-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": 0.025330564007163048,
"advantage/min": -0.9999988079071045,
"advantage/max": 8.083512306213379,
"time/assemble_training_data": 6.207695722579956,
"time/kl_vs_base": 136.1249668598175,
"kl_policy_base": 0.0008551827049814165,
"time/train": 1054.6262123584747,
"time/save_checkpoint": 19.523678302764893,
"time/total": 3049.7484831809998
}
[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