Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step20
ttt-discover-v3-qwen3-8b-erdos-s2-step20 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 19 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 16, 2026
Repo size
354 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 99%
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 19 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 19,
"progress/batch": 19,
"optim/lr": 4e-05,
"progress/done_frac": 0.4,
"puct/buffer_size": 312,
"puct/sampled_size": 8,
"puct/T": 9728,
"puct/scale_last": 0.11907465412874174,
"puct/buffer_value/mean": -0.38559593011280385,
"puct/buffer_value/std": 0.02108334100095046,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.38092534587125826,
"puct/buffer_timestep/mean": 8.743589743589743,
"puct/buffer_timestep/std": 5.632861224571347,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 18.0,
"puct/buffer_construction_len/mean": 79.98717948717949,
"puct/buffer_construction_len/std": 25.299738313100672,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38092624103211536,
"puct/sampled_value/std": 5.015904252371461e-07,
"puct/sampled_value/min": -0.3809269265631607,
"puct/sampled_value/max": -0.38092534587125826,
"puct/sampled_timestep/mean": 18.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 18.0,
"puct/sampled_timestep/max": 18.0,
"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": 2655.8934259414673,
"env/all/ac_tokens_per_turn": 9263.55859375,
"env/all/ob_tokens_per_turn": 1845.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4742942,
"env/all/total_ob_tokens": 944704,
"env/all/time/sampling_mean": 531.8326593767852,
"env/all/time/sampling_max": 799.0678646564484,
"env/all/time/env_step_mean": 524.7851678640582,
"env/all/time/env_step_max": 1900.778261423111,
"env/all/reward/mean": 0.8879899588801996,
"env/all/reward/max": 2.6251862490820996,
"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": 0.8879899588801996,
"env/all/correctness": 0.34375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.38840643837114786,
"env/all/raw_score/min": 0.3809253435247838,
"env/all/raw_score/max": 0.5216919660930394,
"env/all/initial_raw_score": -0.3809262410321154,
"env/all/initial_raw_score/min": -0.3809269265631607,
"env/all/initial_raw_score/max": -0.38092534587125826,
"env/all/msg": "Success; raw_score=0.3809253435247838",
"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 # Use a smaller n to stay within budget\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Generate multiple initial guesses\n initial_guesses = []\n\n # Provided initial guess\n initial_guesses.append(initial_h_values.copy())\n\n # Random binary guess (0/1) with desired integral\n def generate_random_binary(n_points):\n h = np.random.randint(0, 2, size=n_points)\n current_sum = np.sum(h)\n desired_sum = n_points / 2\n delta = desired_sum - current_sum\n if delta > 0:\n for i in range(n_points):\n if delta <= 0:\n break\n if h[i] == 0:\n h[i] = 1\n delta -= 1\n elif delta < 0:\n for i in range(n_points):\n if delta >= 0:\n break\n if h[i] == 1:\n h[i] = 0\n delta += 1\n return h\n\n initial_guesses.append(generate_random_binary(n_points))\n\n # Periodic binary guess with evenly spaced intervals\n def generate_periodic_binary(n_points):\n m = 5 # number of intervals\n h = np.zeros(n_points)\n for i in range(m):\n start = i * (2 / m)\n end = (i + 1) * (2 / m)\n start_idx = int(start / (2 / n_points))\n end_idx = int(end / (2 / n_points))\n for j in range(start_idx, end_idx):\n h[j] = 1.0\n # Adjust to meet sum constraint\n current_sum = np.sum(h)\n desired_sum = n_points / 2\n delta = desired_sum - current_sum\n if delta > 0:\n for i in range(n_points):\n if h[i] == 0 and delta > 0:\n h[i] = 1.0\n delta -= 1\n elif delta < 0:\n for i in range(n_points):\n if h[i] == 1 and delta < 0:\n h[i] = 0.0\n delta += 1\n return h\n\n initial_guesses.append(generate_periodic_binary(n_points))\n\n best_h = None\n best_c5 = float('inf')\n\n for h_initial in initial_guesses:\n h_initial = np.array(h_initial, dtype=np.float64)\n h_initial = np.clip(h_initial, 0.0, 1.0)\n current_sum = np.sum(h_initial)\n desired_sum = n_points / 2\n delta = desired_sum - current_sum\n if delta != 0:\n if delta > 0:\n h_initial[-1] = min(1.0, h_initial[-1] + delta)\n else:\n h_initial[-1] = max(0.0, h_initial[-1] + delta)\n h_initial = np.clip(h_initial, 0.0, 1.0)\n\n # Objective function: max of convolution over k\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 # Constraint: sum of h is n_points / 2\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 # First stage: Basinhopping for global search\n result_bh = basinhopping(\n objective,\n h_initial,\n niter=50, # Reduced from 200\n T=0.2,\n stepsize=0.7,\n minimizer_kwargs={\n 'method': 'SLSQP',\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 # Iterative refinement focusing on top 5 overlapping k's\n for _ in range(300):\n h_array = np.array(best_h_temp, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n top_ks = np.argsort(corr)[::-1][:5] # Pick top 5 k's\n\n # Compute contributions of each point to the top 5 k's\n contributions = np.zeros(n_points)\n for k in top_ks:\n for i in range(n_points):\n if i + k < n_points:\n contributions[i] += h_array[i] * (1 - h_array[i + k])\n else:\n contributions[i] += h_array[i] * 1.0\n sorted_indices = np.argsort(contributions)[::-1]\n top_indices = sorted_indices[:20] # Perturb top 20 contributors\n\n # Perturb and optimize\n h_perturbed = best_h_temp.copy()\n for i in top_indices:\n h_perturbed[i] += np.random.uniform(-0.01, 0.01)\n h_perturbed[i] = np.clip(h_perturbed[i], 0.0, 1.0)\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_perturbed[:-1])))\n\n res = minimize(\n fun=objective,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 500}\n )\n if res.fun < best_c5_temp:\n best_h_temp, best_c5_temp = res.x, res.fun\n\n # Final refinement\n for _ in range(200):\n h_array = np.array(best_h_temp, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n max_k = np.argmax(corr)\n contributions = np.zeros(n_points)\n for i in range(n_points):\n if i + max_k < n_points:\n contributions[i] = h_array[i] * (1 - h_array[i + max_k])\n else:\n contributions[i] = h_array[i] * 1.0\n sorted_indices = np.argsort(contributions)[::-1]\n top_indices = sorted_indices[:20]\n\n h_perturbed = best_h_temp.copy()\n for i in top_indices:\n h_perturbed[i] += np.random.uniform(-0.01, 0.01)\n h_perturbed[i] = np.clip(h_perturbed[i], 0.0, 1.0)\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_perturbed[:-1])))\n\n res = minimize(\n fun=objective,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 500}\n )\n if res.fun < best_c5_temp:\n best_h_temp, best_c5_temp = res.x, res.fun\n\n # Update global best\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": 531.8326593767852,
"env/all/time/policy/min": 194.10930967330933,
"env/all/time/policy/max": 799.0678646564484,
"env/all/time/env_step": 524.7851678640582,
"env/all/time/env_step/min": 0.006904125213623047,
"env/all/time/env_step/max": 1900.778261423111,
"env/all/time/reward_compute": 2.845190465450287e-07,
"env/all/time/reward_compute/min": 2.4586915969848633e-07,
"env/all/time/reward_compute/max": 3.3155083656311035e-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": 0.023654349148273468,
"advantage/min": -0.9999152421951294,
"advantage/max": 3.106889247894287,
"time/assemble_training_data": 8.471657991409302,
"time/kl_vs_base": 138.72320771217346,
"kl_policy_base": 0.0008562975563108921,
"time/train": 1040.2018694877625,
"time/save_checkpoint": 14.01821494102478,
"time/total": 3858.9531598091125
}
[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