Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step10
ttt-discover-v3-qwen3-8b-erdos-s2-step10 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 9 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 16, 2026
Repo size
354 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 99%
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 9 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 9,
"progress/batch": 9,
"optim/lr": 4e-05,
"progress/done_frac": 0.2,
"puct/buffer_size": 152,
"puct/sampled_size": 8,
"puct/T": 4608,
"puct/scale_last": 0.030378839924146306,
"puct/buffer_value/mean": -0.3888771302410919,
"puct/buffer_value/std": 0.026596014902305898,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.3810008138332892,
"puct/buffer_timestep/mean": 3.736842105263158,
"puct/buffer_timestep/std": 2.7499685216027645,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 8.0,
"puct/buffer_construction_len/mean": 64.40789473684211,
"puct/buffer_construction_len/std": 22.31985051802127,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 128.0,
"puct/sampled_value/mean": -0.38103736559859913,
"puct/sampled_value/std": 4.965469206277695e-05,
"puct/sampled_value/min": -0.3811616891740291,
"puct/sampled_value/max": -0.3810008138332892,
"puct/sampled_timestep/mean": 8.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 8.0,
"puct/sampled_timestep/max": 8.0,
"puct/sampled_construction_len/mean": 79.0,
"puct/sampled_construction_len/std": 13.96424004376894,
"puct/sampled_construction_len/min": 64.0,
"puct/sampled_construction_len/max": 100.0,
"time/sampling": 2375.004236936569,
"env/all/ac_tokens_per_turn": 8552.27734375,
"env/all/ob_tokens_per_turn": 1669.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4378766,
"env/all/total_ob_tokens": 854592,
"env/all/time/sampling_mean": 464.77978167822585,
"env/all/time/sampling_max": 660.2872686386108,
"env/all/time/env_step_mean": 406.7369664153084,
"env/all/time/env_step_max": 1733.346664905548,
"env/all/reward/mean": 0.6933225129111186,
"env/all/reward/max": 2.6250235589502973,
"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": 0.6933225129111186,
"env/all/correctness": 0.265625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.38359488603138037,
"env/all/raw_score/min": 0.38094895199708134,
"env/all/raw_score/max": 0.5183051091904262,
"env/all/initial_raw_score": -0.38103736559859913,
"env/all/initial_raw_score/min": -0.3811616891740291,
"env/all/initial_raw_score/max": -0.3810008138332892,
"env/all/msg": "Success; raw_score=0.38162912248658537",
"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 = 200 # Increased resolution for better accuracy\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Generate initial guess with random values satisfying the sum constraint\n h_initial = np.random.uniform(0.0, 1.0, n_points)\n h_initial[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_initial[:-1])))\n\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 max_corr = np.max(corr)\n return max_corr * dx\n\n def constraint_func(h):\n return np.sum(h) - n_points / 2\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Step 1: Use DE with penalty for constraint violation\n def objective_de(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)\n penalty = 1e4 * (np.sum(h_array) - n_points / 2) ** 2\n return max_corr * dx + penalty\n\n result_de = differential_evolution(\n objective_de,\n bounds,\n strategy='best1bin',\n popsize=20,\n maxiter=100,\n tol=1e-6,\n seed=seed\n )\n\n best_h = result_de.x\n best_c5 = result_de.fun\n\n # Step 2: Refine with SLSQP\n res_slsqp = minimize(\n fun=objective,\n x0=best_h,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 2000}\n )\n\n best_h, best_c5 = res_slsqp.x, res_slsqp.fun\n\n # Step 3: Perturbation-based refinement\n for _ in range(50):\n h_perturbed = best_h + np.random.uniform(-0.05, 0.05, size=n_points)\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_perturbed[:-1])))\n\n res_perturbed = minimize(\n fun=objective,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 300}\n )\n\n if res_perturbed.fun < best_c5:\n best_h, best_c5 = res_perturbed.x, res_perturbed.fun\n\n # Final fine-tuning\n for _ in range(100):\n h_perturbed = best_h + np.random.uniform(-0.02, 0.02, size=n_points)\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_perturbed[:-1])))\n\n res_perturbed = minimize(\n fun=objective,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 150}\n )\n\n if res_perturbed.fun < best_c5:\n best_h, best_c5 = res_perturbed.x, res_perturbed.fun\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 464.77978167822585,
"env/all/time/policy/min": 236.0119984149933,
"env/all/time/policy/max": 660.2872686386108,
"env/all/time/env_step": 406.7369664153084,
"env/all/time/env_step/min": 0.0067255496978759766,
"env/all/time/env_step/max": 1733.346664905548,
"env/all/time/reward_compute": 4.98257577419281e-07,
"env/all/time/reward_compute/min": 2.7194619178771973e-07,
"env/all/time/reward_compute/max": 1.4826655387878418e-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": 0.028905415907502174,
"advantage/min": -1.0,
"advantage/max": 4.0232439041137695,
"time/assemble_training_data": 9.138299703598022,
"time/kl_vs_base": 126.67117238044739,
"kl_policy_base": 0.0007217746460810304,
"time/train": 937.2914202213287,
"time/save_checkpoint": 9.808664321899414,
"time/total": 3459.095883846283
}
[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