Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step12
ttt-discover-v3-qwen3-8b-erdos-step12 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 11 (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 11 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 11,
"progress/batch": 11,
"optim/lr": 4e-05,
"progress/done_frac": 0.24,
"puct/buffer_size": 184,
"puct/sampled_size": 8,
"puct/T": 5632,
"puct/scale_last": 0.06766266146164934,
"puct/buffer_value/mean": -0.387840111833764,
"puct/buffer_value/std": 0.026096355481982,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809417650206355,
"puct/buffer_timestep/mean": 4.739130434782608,
"puct/buffer_timestep/std": 3.326015912853092,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 10.0,
"puct/buffer_construction_len/mean": 77.53804347826087,
"puct/buffer_construction_len/std": 16.233308098603825,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38094441534973644,
"puct/sampled_value/std": 3.8973716963281854e-06,
"puct/sampled_value/min": -0.3809511517198728,
"puct/sampled_value/max": -0.3809417650206355,
"puct/sampled_timestep/mean": 10.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 10.0,
"puct/sampled_timestep/max": 10.0,
"puct/sampled_construction_len/mean": 80.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 80.0,
"puct/sampled_construction_len/max": 80.0,
"time/sampling": 1668.6049280166626,
"env/all/ac_tokens_per_turn": 9002.0,
"env/all/ob_tokens_per_turn": 1483.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4609024,
"env/all/total_ob_tokens": 759488,
"env/all/time/sampling_mean": 481.24170217616484,
"env/all/time/sampling_max": 687.5009951591492,
"env/all/time/env_step_mean": 18.830441849771887,
"env/all/time/env_step_max": 1100.1227111816406,
"env/all/reward/mean": 1.3282441336546427,
"env/all/reward/max": 2.6250730849054973,
"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": 1.3282441336546427,
"env/all/correctness": 0.515625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3913306852922984,
"env/all/raw_score/min": 0.38094176482148084,
"env/all/raw_score/max": 1.0,
"env/all/initial_raw_score": -0.38094441534973644,
"env/all/initial_raw_score/min": -0.3809511517198728,
"env/all/initial_raw_score/max": -0.3809417650206355,
"env/all/msg": "RuntimeError: Program execution failed: TypeError: differential_evolution() got an unexpected keyword argument 'fun'",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\nfrom scipy.signal import correlate\n\ndef generate_initial_guesses(n_points, required_sum):\n \"\"\"Generate diverse initial function configurations.\"\"\"\n guesses = []\n\n # Original initial guess\n h = initial_h_values.copy()\n h = h * required_sum / np.sum(h)\n guesses.append(h)\n\n # Perturbed random version\n perturbed = h + np.random.normal(0, 0.05, n_points)\n perturbed = np.clip(perturbed, 0.0, 1.0)\n perturbed = perturbed * required_sum / np.sum(perturbed)\n guesses.append(perturbed)\n\n # Evenly spaced 1s\n h_even = np.zeros(n_points)\n for i in range(n_points):\n if i % 2 == 0:\n h_even[i] = 1\n h_even = h_even * required_sum / np.sum(h_even)\n guesses.append(h_even)\n\n # Random cluster-based guess\n cluster_start = np.random.choice(n_points, size=4, replace=False)\n cluster_end = np.random.choice(n_points, size=4, replace=False)\n h_clusters = np.zeros(n_points)\n for i, j in zip(sorted(cluster_start), sorted(cluster_end)):\n h_clusters[i:j] = 1\n h_clusters = h_clusters * required_sum / np.sum(h_clusters)\n guesses.append(h_clusters)\n\n # Random binary guess\n indices = np.random.choice(n_points, size=required_sum, replace=False)\n h_binary = np.zeros(n_points)\n h_binary[indices] = 1\n guesses.append(h_binary)\n\n # Sine wave guess\n h_sine = 0.5 + 0.5 * np.sin(np.pi * np.arange(n_points) / (2.0))\n h_sine = h_sine * required_sum / np.sum(h_sine)\n guesses.append(h_sine)\n\n # Cosine wave guess\n h_cosine = 0.5 + 0.5 * np.cos(np.pi * np.arange(n_points) / (2.0))\n h_cosine = h_cosine * required_sum / np.sum(h_cosine)\n guesses.append(h_cosine)\n\n # Triangular wave guess\n h_tri = np.minimum(np.arange(n_points)/(n_points/2), np.flipud(np.arange(n_points)/(n_points/2)))\n h_tri = h_tri * required_sum / np.sum(h_tri)\n guesses.append(h_tri)\n\n return guesses\n\ndef evaluate_C5(h, dx):\n \"\"\"Evaluate the overlap integral for all shifts.\"\"\"\n h1 = 1.0 - h\n corr = correlate(h, h1, mode='full')\n max_corr = np.max(corr)\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = int(n_points / 2)\n \n # Define the penalized cost function\n def objective(x):\n c5 = evaluate_C5(x, dx)\n # Penalize for constraint violation\n penalty = 1e3 * (np.sum(x) - required_sum) ** 2\n return c5 + penalty\n \n # Define bounds\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n \n # Run differential evolution\n result = differential_evolution(\n fun=objective,\n bounds=bounds,\n strategy='best1bin',\n popsize=20,\n maxiter=500,\n tol=1e-6,\n mutation=(0.5, 1.0),\n recombination=0.9,\n disp=False\n )\n \n best_h = result.x\n best_c5 = evaluate_C5(best_h, dx)\n \n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 481.24170217616484,
"env/all/time/policy/min": 224.95991969108582,
"env/all/time/policy/max": 687.5009951591492,
"env/all/time/env_step": 18.830441849771887,
"env/all/time/env_step/min": 0.004308462142944336,
"env/all/time/env_step/max": 1100.1227111816406,
"env/all/time/reward_compute": 4.284083843231201e-07,
"env/all/time/reward_compute/min": 2.4586915969848633e-07,
"env/all/time/reward_compute/max": 7.227063179016113e-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": 0.02130313403904438,
"advantage/min": -1.0,
"advantage/max": 8.330735206604004,
"time/assemble_training_data": 7.129070043563843,
"time/kl_vs_base": 127.46492171287537,
"kl_policy_base": 0.0007848198874853551,
"time/train": 972.4235572814941,
"time/save_checkpoint": 11.297631978988647,
"time/total": 2787.9608883857727
}
[2026-07-09T06:24:18+00:00] job=1812624 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:31:44+00:00] job=1812955 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T07:59:00+00:00] job=1813123 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:28:32+00:00] job=1813124 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T09:35:59+00:00] job=1813609 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:45:57+00:00] job=1813622 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes