Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step6
ttt-discover-v3-qwen3-8b-erdos-s2-step6 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 5 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 9, 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 5 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 5,
"progress/batch": 5,
"optim/lr": 4e-05,
"progress/done_frac": 0.12,
"puct/buffer_size": 88,
"puct/sampled_size": 8,
"puct/T": 2560,
"puct/scale_last": 0.03008711523170987,
"puct/buffer_value/mean": -0.39423231670070324,
"puct/buffer_value/std": 0.033952087770159754,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.38129253852572564,
"puct/buffer_timestep/mean": 1.7272727272727273,
"puct/buffer_timestep/std": 1.6006197146962735,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 4.0,
"puct/buffer_construction_len/mean": 62.65909090909091,
"puct/buffer_construction_len/std": 22.392615647511796,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 100.0,
"puct/sampled_value/mean": -0.3813539007005865,
"puct/sampled_value/std": 2.811058665206406e-05,
"puct/sampled_value/min": -0.381396571111547,
"puct/sampled_value/max": -0.38129253852572564,
"puct/sampled_timestep/mean": 4.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 4.0,
"puct/sampled_timestep/max": 4.0,
"puct/sampled_construction_len/mean": 44.875,
"puct/sampled_construction_len/std": 6.2935979375870525,
"puct/sampled_construction_len/min": 40.0,
"puct/sampled_construction_len/max": 53.0,
"time/sampling": 1818.3092260360718,
"env/all/ac_tokens_per_turn": 9077.958984375,
"env/all/ob_tokens_per_turn": 1502.75,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4647915,
"env/all/total_ob_tokens": 769408,
"env/all/time/sampling_mean": 507.17281007627025,
"env/all/time/sampling_max": 736.3447539806366,
"env/all/time/env_step_mean": 91.77245101658627,
"env/all/time/env_step_max": 1100.1203417778015,
"env/all/reward/mean": 0.952704806234203,
"env/all/reward/max": 2.6226617146734745,
"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": 0.952704806234203,
"env/all/correctness": 0.373046875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3937294303846973,
"env/all/raw_score/min": 0.38129201649549543,
"env/all/raw_score/max": 0.5199246943845193,
"env/all/initial_raw_score": -0.3813539007005864,
"env/all/initial_raw_score/min": -0.381396571111547,
"env/all/initial_raw_score/max": -0.38129253852572564,
"env/all/msg": "Success; raw_score=0.38196229354270633",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\n\ndef create_binary_initial(n_points):\n \"\"\"Create a binary function with 1s evenly spaced.\"\"\"\n h = np.zeros(n_points)\n for i in range(n_points // 2):\n h[i] = 1.0\n return h\n\ndef create_symmetric_initial(n_points):\n \"\"\"Create a symmetric function with cosine-weighted values.\"\"\"\n h = np.zeros(n_points)\n for i in range(n_points):\n x = i * 2.0 / n_points - 1.0 # center in [-1, 1]\n # Weighted value to be symmetric\n h[i] = 0.5 * (1.0 + np.cos(np.pi * x))\n # Clip to [0,1] and normalize\n h = np.clip(h, 0.0, 1.0)\n h = h / np.sum(h) * n_points / 2\n return h\n\ndef run(seed=42, budget_s=1000, **kwargs):\n import warnings\n warnings.filterwarnings(\"ignore\", category=RuntimeWarning)\n global h_initial\n h_initial = initial_h_values # Provided by external context\n n_points = len(h_initial)\n dx = 2.0 / n_points\n\n # Generate multiple starting points\n starting_points = [h_initial.copy()]\n starting_points.append(create_binary_initial(n_points))\n starting_points.append(create_symmetric_initial(n_points))\n\n best_h = None\n best_c5 = float('inf')\n seed_counter = seed\n\n for start in starting_points:\n # Define the penalty-adjusted objective function 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 max_corr = np.max(corr) * dx\n # Reduced penalty coefficient for better balance\n penalty = 1e1 * (np.sum(h_array) - n_points / 2) ** 2\n return max_corr + penalty\n\n np.random.seed(seed_counter)\n seed_counter += 1\n\n # Run differential evolution\n result_de = differential_evolution(\n objective_with_penalty,\n bounds=[(0.0, 1.0) for _ in range(n_points)],\n strategy='rand1bin',\n maxiter=100,\n popsize=50,\n tol=1e-5,\n mutation=(0.6, 0.9),\n recombination=0.8,\n seed=seed_counter\n )\n\n # Extract DE result\n h_de = result_de.x.copy()\n # Enforce the integral constraint\n sum_h = np.sum(h_de)\n sum_prev = np.sum(h_de[:-1])\n h_de[-1] = max(0.0, min(1.0, n_points / 2 - sum_prev))\n h_de[-1] = np.clip(h_de[-1], 0.0, 1.0)\n while np.abs(np.sum(h_de) - n_points / 2) > 1e-6:\n sum_prev = np.sum(h_de[:-1])\n h_de[-1] = max(0.0, min(1.0, n_points / 2 - sum_prev))\n\n # Objective function for SLSQP\n def objective_slsqp(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 # Constraints for SLSQP\n constraints = [\n {'type': 'eq', 'fun': lambda h: np.sum(h) - n_points / 2}\n ]\n\n # Bounds for SLSQP\n bounds_slsqp = [(0.0, 1.0) for _ in range(n_points)]\n\n # Refine the DE solution with SLSQP\n res_slsqp = minimize(\n fun=objective_slsqp,\n x0=h_de,\n method='SLSQP',\n bounds=bounds_slsqp,\n constraints=constraints,\n options={'ftol': 1e-9, 'maxiter': 200}\n )\n\n # Perturbation step to escape local optima\n for _ in range(5):\n h_perturbed = res_slsqp.x + np.random.uniform(-0.05, 0.05, size=n_points)\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n sum_perturbed = np.sum(h_perturbed[:-1])\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - sum_perturbed))\n res_perturbed = minimize(\n fun=objective_slsqp,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds_slsqp,\n constraints=constraints,\n options={'ftol': 1e-9, 'maxiter': 100}\n )\n if res_perturbed.fun < res_slsqp.fun:\n res_slsqp = res_perturbed\n\n # Final refinement step\n for _ in range(5):\n h_perturbed = res_slsqp.x + np.random.uniform(-0.01, 0.01, size=n_points)\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n sum_perturbed = np.sum(h_perturbed[:-1])\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - sum_perturbed))\n res_perturbed = minimize(\n fun=objective_slsqp,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds_slsqp,\n constraints=constraints,\n options={'ftol': 1e-9, 'maxiter': 50}\n )\n if res_perturbed.fun < res_slsqp.fun:\n res_slsqp = res_perturbed\n\n # Update best solution\n if res_slsqp.fun < best_c5:\n best_c5 = res_slsqp.fun\n best_h = res_slsqp.x\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 507.17281007627025,
"env/all/time/policy/min": 241.3972990512848,
"env/all/time/policy/max": 736.3447539806366,
"env/all/time/env_step": 91.77245101658627,
"env/all/time/env_step/min": 0.005120754241943359,
"env/all/time/env_step/max": 1100.1203417778015,
"env/all/time/reward_compute": 4.735775291919708e-07,
"env/all/time/reward_compute/min": 2.3096799850463867e-07,
"env/all/time/reward_compute/max": 1.1473894119262695e-06,
"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": 0.022953644394874573,
"advantage/min": -1.0,
"advantage/max": 2.823169708251953,
"time/assemble_training_data": 9.459734201431274,
"time/kl_vs_base": 130.0426890850067,
"kl_policy_base": 0.0006039614672772586,
"time/train": 982.777768611908,
"time/save_checkpoint": 7.357166767120361,
"time/total": 2949.6335096359253
}
[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