Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step23
ttt-discover-v3-qwen3-8b-erdos-s2-step23 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 22 (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 22 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 22,
"progress/batch": 22,
"optim/lr": 4e-05,
"progress/done_frac": 0.46,
"puct/buffer_size": 360,
"puct/sampled_size": 8,
"puct/T": 11264,
"puct/scale_last": 0.11907469918514058,
"puct/buffer_value/mean": -0.38497622886881766,
"puct/buffer_value/std": 0.019691025198411236,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.3809253008148594,
"puct/buffer_timestep/mean": 10.244444444444444,
"puct/buffer_timestep/std": 6.4983948635389295,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 21.0,
"puct/buffer_construction_len/mean": 82.65555555555555,
"puct/buffer_construction_len/std": 24.515600978928184,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.3809257093125884,
"puct/sampled_value/std": 3.8042416682611387e-07,
"puct/sampled_value/min": -0.3809261277347607,
"puct/sampled_value/max": -0.3809253008148594,
"puct/sampled_timestep/mean": 21.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 21.0,
"puct/sampled_timestep/max": 21.0,
"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": 2170.9865894317627,
"env/all/ac_tokens_per_turn": 9364.015625,
"env/all/ob_tokens_per_turn": 1763.25,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4794376,
"env/all/total_ob_tokens": 902784,
"env/all/time/sampling_mean": 526.5012073828839,
"env/all/time/sampling_max": 789.6645851135254,
"env/all/time/env_step_mean": 209.4978067255579,
"env/all/time/env_step_max": 1384.499998807907,
"env/all/reward/mean": 0.8871079552311623,
"env/all/reward/max": 2.62518653919221,
"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": 0.8871079552311623,
"env/all/correctness": 0.34765625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.39406448171289576,
"env/all/raw_score/min": 0.3809253014286148,
"env/all/raw_score/max": 0.5049738746817931,
"env/all/initial_raw_score": -0.3809257093125884,
"env/all/initial_raw_score/min": -0.3809261277347607,
"env/all/initial_raw_score/max": -0.3809253008148594,
"env/all/msg": "RuntimeError: Program execution failed: TypeError: _minimize_trustregion_constr() got an unexpected keyword argument 'ftol'",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import basinhopping, minimize\nfrom scipy.optimize.optimize import OptimizeResult\nfrom scipy.optimize import differential_evolution\n\nimport math\n\ndef run(seed=42, budget_s=1000, **kwargs):\n # Increase resolution for better function representation\n n_points = 200\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Generate a better initial guess (binary function)\n num_ones = n_points // 2\n initial_h = np.zeros(n_points)\n indices = np.random.choice(n_points, num_ones, replace=False)\n initial_h[indices] = 1.0\n initial_h = np.clip(initial_h, 0.0, 1.0)\n # Normalize to ensure sum is n_points / 2\n current_sum = np.sum(initial_h)\n if abs(current_sum - n_points / 2) > 1e-6:\n delta = (n_points / 2) - current_sum\n # Adjust last element\n initial_h[-1] += delta\n initial_h = np.clip(initial_h, 0.0, 1.0)\n\n best_h = initial_h.copy()\n best_c5 = float('inf')\n\n def objective(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 # Step 1: Basinhopping for global optimization\n result_bh = basinhopping(\n objective,\n initial_h,\n niter=300, # More iterations for better exploration\n T=0.1, # Lower temperature to focus on good candidates\n stepsize=0.5,\n minimizer_kwargs={\n 'method': 'SLSQP', # For constrained optimization\n 'bounds': bounds,\n 'constraints': [{'type': 'eq', 'fun': constraint_func}],\n 'options': {'ftol': 1e-10, 'maxiter': 500, 'disp': False}\n },\n seed=seed\n )\n\n best_h_temp = result_bh.x\n best_c5_temp = result_bh.fun\n\n # Step 2: Trust-Constrained Local Refinement\n # Use a better local optimizer for refinement\n result_l = minimize(\n fun=objective,\n x0=best_h_temp,\n method='trust-constr',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 500}\n )\n\n best_h_temp = result_l.x\n best_c5_temp = result_l.fun\n\n # Step 3: Add another refinement call with different settings\n # for more robustness\n result_l2 = minimize(\n fun=objective,\n x0=best_h_temp,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 500}\n )\n\n if result_l2.fun < best_c5_temp:\n best_h_temp = result_l2.x\n best_c5_temp = result_l2.fun\n\n # Step 4: Compare to existing best and update\n if best_c5_temp < best_c5:\n best_h = best_h_temp\n best_c5 = best_c5_temp\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 526.5012073828839,
"env/all/time/policy/min": 239.73344588279724,
"env/all/time/policy/max": 789.6645851135254,
"env/all/time/env_step": 209.4978067255579,
"env/all/time/env_step/min": 0.006326198577880859,
"env/all/time/env_step/max": 1384.499998807907,
"env/all/time/reward_compute": 3.688037395477295e-07,
"env/all/time/reward_compute/min": 2.4586915969848633e-07,
"env/all/time/reward_compute/max": 7.636845111846924e-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": 0.023564469069242477,
"advantage/min": -0.9765719175338745,
"advantage/max": 2.8747024536132812,
"time/assemble_training_data": 9.356364965438843,
"time/kl_vs_base": 137.93403577804565,
"kl_policy_base": 0.0008740578778088093,
"time/train": 1042.7014157772064,
"time/save_checkpoint": 10.92229700088501,
"time/total": 3373.1738402843475
}
[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