Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step2
ttt-discover-v3-qwen3-8b-erdos-s2-step2 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 1 (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 1 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 1,
"progress/batch": 1,
"optim/lr": 4e-05,
"progress/done_frac": 0.04,
"puct/buffer_size": 24,
"puct/sampled_size": 8,
"puct/T": 512,
"puct/scale_last": 0.028795268218252956,
"puct/buffer_value/mean": -0.4273486942870435,
"puct/buffer_value/std": 0.052135821437967814,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.38258438553918256,
"puct/buffer_timestep/mean": -0.3333333333333333,
"puct/buffer_timestep/std": 0.4714045207910317,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 0.0,
"puct/buffer_construction_len/mean": 65.75,
"puct/buffer_construction_len/std": 23.13952246698276,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 99.0,
"puct/sampled_value/mean": -0.38445518084447494,
"puct/sampled_value/std": 0.0017123584518135971,
"puct/sampled_value/min": -0.3880241467584787,
"puct/sampled_value/max": -0.38258438553918256,
"puct/sampled_timestep/mean": 0.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 0.0,
"puct/sampled_timestep/max": 0.0,
"puct/sampled_construction_len/mean": 63.0,
"puct/sampled_construction_len/std": 23.979157616563597,
"puct/sampled_construction_len/min": 40.0,
"puct/sampled_construction_len/max": 99.0,
"time/sampling": 1219.733558177948,
"env/all/ac_tokens_per_turn": 9317.1953125,
"env/all/ob_tokens_per_turn": 1111.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4770404,
"env/all/total_ob_tokens": 569024,
"env/all/time/sampling_mean": 513.5577502613887,
"env/all/time/sampling_max": 741.8569021224976,
"env/all/time/env_step_mean": 17.044354510027915,
"env/all/time/env_step_max": 568.1773192882538,
"env/all/reward/mean": 0.9592252508516964,
"env/all/reward/max": 2.619911340945217,
"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": 0.9592252508516964,
"env/all/correctness": 0.38671875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.4069073117320764,
"env/all/raw_score/min": 0.38169229552634576,
"env/all/raw_score/max": 0.5338414415921754,
"env/all/initial_raw_score": -0.38445518084447494,
"env/all/initial_raw_score/min": -0.3880241467584787,
"env/all/initial_raw_score/max": -0.38258438553918256,
"env/all/msg": "Success; raw_score=0.3825843855392442",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\n\ndef run(seed=42, budget_s=1000, **kwargs):\n # Initialize with the provided initial function\n h = np.copy(initial_h_values)\n n_points = len(h)\n dx = 2.0 / n_points\n\n # Define the objective function with a penalty for constraint violation\n def objective(h_vec):\n correlate_result = np.correlate(h_vec, 1.0 - h_vec, mode='full') * dx\n max_correlate = np.max(correlate_result)\n constraint_penalty = 1e5 * (np.sum(h_vec) - n_points / 2) ** 2\n return max_correlate + constraint_penalty\n\n # Define bounds for each element\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Generate a few random starting points\n start_points = [h]\n np.random.seed(seed)\n\n # Generate a random starting point with sum(h) = n_points/2\n random_start = np.random.rand(n_points)\n scaling = (n_points / 2) / np.sum(random_start)\n random_start *= scaling\n start_points.append(random_start)\n\n # Additional starting patterns for diversification\n # For n_points even, try a simple alternating 1 and 0 pattern\n if n_points % 2 == 0:\n alt_pattern = np.zeros(n_points)\n alt_pattern[::2] = 1.0\n alt_pattern = alt_pattern * (n_points / 2) / np.sum(alt_pattern)\n start_points.append(alt_pattern)\n else:\n alt_pattern = np.zeros(n_points)\n alt_pattern[::2] = 1.0\n alt_pattern[-1] = 0.5 # adjust for odd n_points\n alt_pattern = alt_pattern * (n_points / 2) / np.sum(alt_pattern)\n start_points.append(alt_pattern)\n\n # Run differential evolution with multiple starting points\n best_h = initial_h_values\n best_c5 = objective(best_h)\n\n for start in start_points:\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n maxiter=100,\n popsize=15,\n tol=1e-6,\n mutation=(0.5, 1),\n recombination=0.9,\n polish=True,\n seed=seed,\n args=()\n )\n current_c5 = result.fun\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_h = result.x\n\n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 513.5577502613887,
"env/all/time/policy/min": 244.57981395721436,
"env/all/time/policy/max": 741.8569021224976,
"env/all/time/env_step": 17.044354510027915,
"env/all/time/env_step/min": 0.005362749099731445,
"env/all/time/env_step/max": 568.1773192882538,
"env/all/time/reward_compute": 3.855675458908081e-07,
"env/all/time/reward_compute/min": 2.123415470123291e-07,
"env/all/time/reward_compute/max": 6.258487701416016e-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": 0.024321280419826508,
"advantage/min": -1.0,
"advantage/max": 3.3898229598999023,
"time/assemble_training_data": 8.379260778427124,
"time/kl_vs_base": 123.18043637275696,
"kl_policy_base": 0.0005257753073237836,
"time/train": 964.1076040267944,
"time/save_checkpoint": 8.548235654830933,
"time/total": 2327.278235912323
}
[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