Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step17
ttt-discover-v3-qwen3-8b-erdos-s2-step17 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 16 (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 16 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 16,
"progress/batch": 16,
"optim/lr": 4e-05,
"progress/done_frac": 0.34,
"puct/buffer_size": 264,
"puct/sampled_size": 8,
"puct/T": 8192,
"puct/scale_last": 0.11907160149367257,
"puct/buffer_value/mean": -0.38643656460656417,
"puct/buffer_value/std": 0.02281925859754162,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.3809283985063274,
"puct/buffer_timestep/mean": 7.242424242424242,
"puct/buffer_timestep/std": 4.767505561041729,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 15.0,
"puct/buffer_construction_len/mean": 76.34848484848484,
"puct/buffer_construction_len/std": 25.891962262940364,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.3809341870361107,
"puct/sampled_value/std": 3.8780399155346955e-06,
"puct/sampled_value/min": -0.38093798532647,
"puct/sampled_value/max": -0.3809283985063274,
"puct/sampled_timestep/mean": 15.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 15.0,
"puct/sampled_timestep/max": 15.0,
"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": 2625.454779624939,
"env/all/ac_tokens_per_turn": 9369.203125,
"env/all/ob_tokens_per_turn": 1605.875,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4797032,
"env/all/total_ob_tokens": 822208,
"env/all/time/sampling_mean": 532.3166689588688,
"env/all/time/sampling_max": 759.1089856624603,
"env/all/time/env_step_mean": 598.9619163214229,
"env/all/time/env_step_max": 1867.438945531845,
"env/all/reward/mean": 0.7568959691242925,
"env/all/reward/max": 2.625175182111452,
"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": 0.7568959691242925,
"env/all/correctness": 0.294921875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3917485805695538,
"env/all/raw_score/min": 0.3809269493946454,
"env/all/raw_score/max": 0.5741463707310682,
"env/all/initial_raw_score": -0.3809341870361107,
"env/all/initial_raw_score/min": -0.38093798532647,
"env/all/initial_raw_score/max": -0.3809283985063274,
"env/all/msg": "RuntimeError: Program execution failed: RuntimeError: The map-like callable must be of the form f(func, iterable), returning a sequence of numbers the same length as 'iterable'",
"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 = 100 # Optimized for speed and accuracy\n dx = 2.0 / n_points\n \n np.random.seed(seed)\n \n # Initial guess based on known construction\n h_initial = initial_h_values.copy()\n # Ensure sum constraint\n h_initial[-1] = (n_points / 2.0) - np.sum(h_initial[:-1])\n h_initial[-1] = np.clip(h_initial[-1], 0.0, 1.0)\n \n def objective_de(h):\n h_array = np.array(h, dtype=np.float64)\n sum_h = np.sum(h_array[:-1])\n h_array[-1] = (n_points / 2.0) - sum_h\n h_array[-1] = np.clip(h_array[-1], 0.0, 1.0)\n \n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n return np.max(corr) * dx\n \n bounds = [(0.0, 1.0) for _ in range(n_points)]\n \n # Global optimization with differential evolution\n result_de = differential_evolution(\n objective_de,\n bounds,\n strategy='best1bin',\n maxiter=100,\n popsize=20,\n tol=1e-6,\n mutation=(0.5, 1.0),\n recombination=0.8,\n disp=False,\n vectorized=True\n )\n \n best_h = result_de.x\n best_c5 = result_de.fun\n\n # Local refinement with targeted perturbation\n for _ in range(150):\n h_perturbed = best_h.copy()\n h_array = np.array(best_h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n max_k = np.argmax(corr)\n \n # Identify contributing indices\n max_contrib_indices = np.argsort(np.abs(corr[max_k]))[-20:] # Select top contributors\n\n for i in max_contrib_indices:\n if i >= n_points - 5: # Avoid perturbation of last element\n continue\n h_perturbed[i] += np.random.uniform(-0.02, 0.02)\n h_perturbed[i] = np.clip(h_perturbed[i], 0.0, 1.0)\n \n # Ensure sum constraint\n h_perturbed[-1] = (n_points / 2.0) - np.sum(h_perturbed[:-1])\n h_perturbed[-1] = np.clip(h_perturbed[-1], 0.0, 1.0)\n \n res = minimize(\n fun=objective_de,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': lambda h: np.sum(h) - n_points / 2}],\n options={'ftol': 1e-10, 'maxiter': 100}\n )\n \n if res.fun < best_c5:\n best_h, best_c5 = res.x, res.fun\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 532.3166689588688,
"env/all/time/policy/min": 234.11358952522278,
"env/all/time/policy/max": 759.1089856624603,
"env/all/time/env_step": 598.9619163214229,
"env/all/time/env_step/min": 0.006444215774536133,
"env/all/time/env_step/max": 1867.438945531845,
"env/all/time/reward_compute": 2.8079375624656677e-07,
"env/all/time/reward_compute/min": 1.7881393432617188e-07,
"env/all/time/reward_compute/max": 4.470348358154297e-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": 0.025810394436120987,
"advantage/min": -0.8575319051742554,
"advantage/max": 2.8747713565826416,
"time/assemble_training_data": 8.489389181137085,
"time/kl_vs_base": 131.6459903717041,
"kl_policy_base": 0.0008315334562212229,
"time/train": 1025.6474335193634,
"time/save_checkpoint": 8.310502767562866,
"time/total": 3801.122829914093
}
[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