Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step42
ttt-discover-v3-qwen3-8b-erdos-s2-step42 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 41 (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 41 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 41,
"progress/batch": 41,
"optim/lr": 4e-05,
"progress/done_frac": 0.84,
"puct/buffer_size": 664,
"puct/sampled_size": 8,
"puct/T": 20992,
"puct/scale_last": 0.27910482463396785,
"puct/buffer_value/mean": -0.38412799949921367,
"puct/buffer_value/std": 0.019367709187158723,
"puct/buffer_value/min": -0.66,
"puct/buffer_value/max": -0.3808951753660322,
"puct/buffer_timestep/mean": 19.746987951807228,
"puct/buffer_timestep/std": 11.981756587728098,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 40.0,
"puct/buffer_construction_len/mean": 96.37951807228916,
"puct/buffer_construction_len/std": 33.373578247903694,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38090889351271684,
"puct/sampled_value/std": 1.0216074638911812e-05,
"puct/sampled_value/min": -0.3809227955826791,
"puct/sampled_value/max": -0.3808951753660322,
"puct/sampled_timestep/mean": 40.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 40.0,
"puct/sampled_timestep/max": 40.0,
"puct/sampled_construction_len/mean": 162.5,
"puct/sampled_construction_len/std": 48.41229182759271,
"puct/sampled_construction_len/min": 100.0,
"puct/sampled_construction_len/max": 200.0,
"time/sampling": 2242.8441817760468,
"env/all/ac_tokens_per_turn": 9801.169921875,
"env/all/ob_tokens_per_turn": 1447.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 5018199,
"env/all/total_ob_tokens": 740928,
"env/all/time/sampling_mean": 570.1751642157324,
"env/all/time/sampling_max": 823.3196520805359,
"env/all/time/env_step_mean": 251.3702168194577,
"env/all/time/env_step_max": 1419.123554468155,
"env/all/reward/mean": 0.5625574524721426,
"env/all/reward/max": 2.625432783990787,
"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": 0.5625574524721426,
"env/all/correctness": 0.2421875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.44372328629493607,
"env/all/raw_score/min": 0.3808895736517859,
"env/all/raw_score/max": 1.0,
"env/all/initial_raw_score": -0.3809088935127168,
"env/all/initial_raw_score/min": -0.3809227955826791,
"env/all/initial_raw_score/max": -0.3808951753660322,
"env/all/msg": "After normalization, h(x) is not in [0, 1]. Range: [0.0, 1.0000000000001241]",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n n_points = 200 # Good balance of accuracy and speed\n dx = 2.0 / n_points\n desired_sum = n_points / 2 # Ensures \u222b\u2080\u00b2 h(x) dx = 1\n\n # Generate multiple initial guesses\n initial_guesses = []\n\n if 'initial_h_values' in globals():\n initial_h = np.array(initial_h_values)\n # Generate 10 perturbations from the known initial guess\n for _ in range(10):\n perturbed = initial_h + np.random.normal(0, 0.05, size=n_points)\n perturbed = np.clip(perturbed, 0.0, 1.0)\n initial_guesses.append(perturbed)\n else:\n # Fallback to random guesses\n for _ in range(10):\n initial_guesses.append(np.random.rand(n_points))\n\n # Adjust each guess to meet the integral constraint\n adjusted_guesses = []\n for h in initial_guesses:\n current_sum = np.sum(h)\n delta = desired_sum - current_sum\n if abs(delta) > 1e-5:\n h = h + delta / n_points\n h = np.clip(h, 0.0, 1.0)\n adjusted_guesses.append(h)\n\n best_t = 1.0\n best_h = np.zeros(n_points)\n\n for initial_h in adjusted_guesses:\n # Define the optimization objective: minimize t\n def objective(x):\n return x[-1]\n\n def constraint_sum(x):\n return np.sum(x[:-1]) - desired_sum\n\n def constraints_func(x):\n h = x[:-1]\n t = x[-1]\n overlaps = []\n\n # Compute overlap for delta from 1 to 10 (only critical shifts)\n for delta in range(1, 11): # 10 shifts\n len_h = n_points - delta\n if len_h > 0:\n # Positive shift\n overlap_pos = np.sum(h[:len_h] * (1 - h[delta:])) * dx\n overlaps.append(overlap_pos)\n # Negative shift\n overlap_neg = np.sum(h[delta:] * (1 - h[:len_h])) * dx\n overlaps.append(overlap_neg)\n\n return [t - overlap for overlap in overlaps]\n\n # Define optimization bounds and constraints\n cons = [\n {'type': 'eq', 'fun': constraint_sum},\n {'type': 'ineq', 'fun': constraints_func}\n ]\n\n bounds = [(0.0, 1.0) for _ in range(n_points)] + [(0.0, np.inf)]\n\n try:\n result = minimize(\n fun=objective,\n x0=np.concatenate((initial_h, [0.5])),\n bounds=bounds,\n constraints=cons,\n method='SLSQP',\n options={\n 'ftol': 1e-7, # Tighter tolerance for better convergence\n 'maxiter': 1000,\n 'disp': False,\n }\n )\n if result.success:\n current_t = result.x[-1]\n if current_t < best_t:\n best_t = current_t\n best_h = result.x[:-1]\n except:\n pass # Skip if optimization fails\n\n return best_h, best_t, n_points\n```",
"env/all/time/policy": 570.1751642157324,
"env/all/time/policy/min": 249.3530502319336,
"env/all/time/policy/max": 823.3196520805359,
"env/all/time/env_step": 251.3702168194577,
"env/all/time/env_step/min": 0.00762939453125,
"env/all/time/env_step/max": 1419.123554468155,
"env/all/time/reward_compute": 3.2549723982810974e-07,
"env/all/time/reward_compute/min": 1.9371509552001953e-07,
"env/all/time/reward_compute/max": 5.476176738739014e-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": 0.03397005796432495,
"advantage/min": -0.9788300395011902,
"advantage/max": 7.789606094360352,
"time/assemble_training_data": 10.132380723953247,
"time/kl_vs_base": 134.5347101688385,
"kl_policy_base": 0.0008978196419775486,
"time/train": 1052.2887399196625,
"time/save_checkpoint": 18.27177119255066,
"time/total": 3459.383884191513
}
[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