Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step19
ttt-discover-v3-qwen3-8b-erdos-s2-step19 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 18 (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 18 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 18,
"progress/batch": 18,
"optim/lr": 4e-05,
"progress/done_frac": 0.38,
"puct/buffer_size": 296,
"puct/sampled_size": 8,
"puct/T": 9216,
"puct/scale_last": 0.11907425591611298,
"puct/buffer_value/mean": -0.385848276572736,
"puct/buffer_value/std": 0.021616959826705966,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.380925744083887,
"puct/buffer_timestep/mean": 8.243243243243244,
"puct/buffer_timestep/std": 5.344385310420472,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 17.0,
"puct/buffer_construction_len/mean": 78.9054054054054,
"puct/buffer_construction_len/std": 25.53146914549843,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38092731378863204,
"puct/sampled_value/std": 1.3970422918030803e-06,
"puct/sampled_value/min": -0.3809303475405988,
"puct/sampled_value/max": -0.380925744083887,
"puct/sampled_timestep/mean": 17.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 17.0,
"puct/sampled_timestep/max": 17.0,
"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": 2655.130049228668,
"env/all/ac_tokens_per_turn": 9107.5,
"env/all/ob_tokens_per_turn": 1938.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4663040,
"env/all/total_ob_tokens": 992448,
"env/all/time/sampling_mean": 522.7265860633925,
"env/all/time/sampling_max": 754.6719038486481,
"env/all/time/env_step_mean": 596.7908811098896,
"env/all/time/env_step_max": 1919.2610402107239,
"env/all/reward/mean": 0.7974774620950027,
"env/all/reward/max": 2.6251862329111297,
"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": 0.7974774620950027,
"env/all/correctness": 0.30859375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3886198037964496,
"env/all/raw_score/min": 0.38092534587125826,
"env/all/raw_score/max": 0.6116966660008447,
"env/all/initial_raw_score": -0.3809273137886321,
"env/all/initial_raw_score/min": -0.3809303475405988,
"env/all/initial_raw_score/max": -0.380925744083887,
"env/all/msg": "Evaluation timed out after 1100 seconds.",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import basinhopping, minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 100 # Optimal for balance between resolution and speed\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Generate a random initial guess that respects the integral constraint\n h_initial = np.random.uniform(0.0, 1.0, size=n_points)\n total = np.sum(h_initial)\n h_initial = h_initial * (n_points / 2) / total\n h_initial = np.clip(h_initial, 0.0, 1.0)\n\n def objective_total(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.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 broader search\n result_bh = basinhopping(\n objective_total,\n h_initial,\n niter=900, # More exploration iterations\n T=0.25, # Slightly higher temperature for exploration\n stepsize=0.8,\n minimizer_kwargs={\n 'method': 'SLSQP',\n 'bounds': bounds,\n 'constraints': [{'type': 'eq', 'fun': constraint_func}],\n 'options': {'ftol': 1e-10, 'maxiter': 2000, 'disp': False}\n },\n seed=seed\n )\n\n best_h = result_bh.x\n best_c5 = result_bh.fun\n\n # Local refinement using gradient-based optimization on the worst k\n def objective_for_k(h, max_k):\n h_arr = np.array(h)\n integral = 0.0\n for i in range(n_points):\n if i + max_k < n_points:\n integral += h_arr[i] * (1 - h_arr[i + max_k])\n else:\n integral += h_arr[i] * 1.0\n return integral * dx\n\n # Gradient of the objective for a specific k\n def gradient_for_k(max_k, h):\n grad = np.zeros(n_points)\n for i in range(n_points):\n if i + max_k < n_points:\n grad[i] = (1 - h[i + max_k]) * dx\n else:\n grad[i] = 1.0 * dx\n return grad\n\n # Refinement loop with guided optimization\n for _ in range(800):\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 # Compute gradient and target objective\n grad = gradient_for_k(max_k, h_array)\n res = minimize(\n fun=lambda h: objective_for_k(h, max_k),\n x0=best_h,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n jac=lambda h: gradient_for_k(max_k, h),\n options={'ftol': 1e-10, 'maxiter': 500}\n )\n\n if res.fun < best_c5:\n best_h, best_c5 = res.x, res.fun\n\n # Final refinement pass\n for _ in range(500):\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 grad = gradient_for_k(max_k, h_array)\n res = minimize(\n fun=lambda h: objective_for_k(h, max_k),\n x0=best_h,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n jac=lambda h: gradient_for_k(max_k, h),\n options={'ftol': 1e-10, 'maxiter': 500}\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": 522.7265860633925,
"env/all/time/policy/min": 203.68966388702393,
"env/all/time/policy/max": 754.6719038486481,
"env/all/time/env_step": 596.7908811098896,
"env/all/time/env_step/min": 0.0062639713287353516,
"env/all/time/env_step/max": 1919.2610402107239,
"env/all/time/reward_compute": 3.348104655742645e-07,
"env/all/time/reward_compute/min": 2.123415470123291e-07,
"env/all/time/reward_compute/max": 4.5821070671081543e-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": 0.0254264697432518,
"advantage/min": -1.0,
"advantage/max": 2.660270929336548,
"time/assemble_training_data": 7.928828477859497,
"time/kl_vs_base": 131.11860060691833,
"kl_policy_base": 0.000898414698895067,
"time/train": 1029.910127401352,
"time/save_checkpoint": 12.087936878204346,
"time/total": 3837.4844279289246
}
[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