Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step14
ttt-discover-v3-qwen3-8b-erdos-s2-step14 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 13 (0-indexed). Strict upstream eval parity: 1100s har…
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 13 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 13,
"progress/batch": 13,
"optim/lr": 4e-05,
"progress/done_frac": 0.28,
"puct/buffer_size": 216,
"puct/sampled_size": 8,
"puct/T": 6656,
"puct/scale_last": 0.030441576396330328,
"puct/buffer_value/mean": -0.38655157684927605,
"puct/buffer_value/std": 0.022596813681451226,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.3809380773611052,
"puct/buffer_timestep/mean": 5.7407407407407405,
"puct/buffer_timestep/std": 3.9024455067027692,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 12.0,
"puct/buffer_construction_len/mean": 70.86111111111111,
"puct/buffer_construction_len/std": 24.415750204011577,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38094301326054436,
"puct/sampled_value/std": 2.850434097813124e-06,
"puct/sampled_value/min": -0.380945595825138,
"puct/sampled_value/max": -0.3809380773611052,
"puct/sampled_timestep/mean": 12.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 12.0,
"puct/sampled_timestep/max": 12.0,
"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": 2890.034734249115,
"env/all/ac_tokens_per_turn": 9504.552734375,
"env/all/ob_tokens_per_turn": 1571.875,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4866331,
"env/all/total_ob_tokens": 804800,
"env/all/time/sampling_mean": 537.5294812787324,
"env/all/time/sampling_max": 777.8540315628052,
"env/all/time/env_step_mean": 720.5417699990794,
"env/all/time/env_step_max": 2102.3101205825806,
"env/all/reward/mean": 0.6587806278491475,
"env/all/reward/max": 2.625098895303768,
"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": 0.6587806278491475,
"env/all/correctness": 0.255859375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3902250517342659,
"env/all/raw_score/min": 0.3809380193401416,
"env/all/raw_score/max": 0.5279588336858922,
"env/all/initial_raw_score": -0.38094301326054436,
"env/all/initial_raw_score/min": -0.380945595825138,
"env/all/initial_raw_score/max": -0.3809380773611052,
"env/all/msg": "RuntimeError: Program execution failed: ValueError: `constraint` of an unknown type is passed.",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\nfrom scipy.signal import correlate\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 200\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Initialize h with initial_h_values scaled to meet the integral constraint\n h_initial = initial_h_values.copy()\n current_sum = np.sum(h_initial)\n scale_factor = (n_points / 2) / current_sum\n h_initial = np.clip(h_initial * scale_factor, 0.0, 1.0)\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 = correlate(h_array, 1 - h_array, mode='full')\n return np.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 # Global optimization with differential evolution\n result_de = differential_evolution(\n objective,\n bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n strategy='best1bin',\n popsize=15,\n maxiter=250,\n mutation=(0.5, 1),\n recombination=0.5,\n tol=1e-10,\n seed=seed\n )\n\n best_h = result_de.x\n best_c5 = result_de.fun\n\n # Refinement: focus on areas of high overlap\n for _ in range(100): # Reduce number of steps to avoid slowdown\n h_perturbed = best_h.copy()\n corr = correlate(h_perturbed, 1 - h_perturbed, mode='full')\n max_corr_idx = np.argmax(corr)\n shift_k = max_corr_idx - (n_points - 1)\n\n # Map shift to index range for perturbation\n if shift_k >= 0:\n start = max(0, shift_k)\n end = min(n_points, shift_k + n_points)\n else:\n start = max(0, n_points + shift_k)\n end = n_points\n\n # Perturb the most relevant area\n perturbation = np.random.uniform(-0.005, 0.005, size=end - start)\n h_perturbed[start:end] += perturbation\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n\n # Ensure the sum constraint is maintained\n last_val = max(0.0, min(1.0, n_points / 2 - np.sum(h_perturbed[:-1])))\n h_perturbed[-1] = last_val\n\n # Local minimization with L-BFGS-B\n res_perturbed = minimize(\n fun=objective,\n x0=h_perturbed,\n method='L-BFGS-B',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 100}\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": 537.5294812787324,
"env/all/time/policy/min": 198.2390491962433,
"env/all/time/policy/max": 777.8540315628052,
"env/all/time/env_step": 720.5417699990794,
"env/all/time/env_step/min": 0.007097482681274414,
"env/all/time/env_step/max": 2102.3101205825806,
"env/all/time/reward_compute": 5.220063030719757e-07,
"env/all/time/reward_compute/min": 2.3096799850463867e-07,
"env/all/time/reward_compute/max": 1.4081597328186035e-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": 0.029761798679828644,
"advantage/min": -0.9331046342849731,
"advantage/max": 5.546158313751221,
"time/assemble_training_data": 8.596660137176514,
"time/kl_vs_base": 132.70002579689026,
"kl_policy_base": 0.0007857091259211302,
"time/train": 1039.3885660171509,
"time/save_checkpoint": 10.425339698791504,
"time/total": 4082.3297386169434
}
[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