Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step40
ttt-discover-v3-qwen3-8b-erdos-step40 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 39 (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 39 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 39,
"progress/batch": 39,
"optim/lr": 4e-05,
"progress/done_frac": 0.8,
"puct/buffer_size": 630,
"puct/sampled_size": 8,
"puct/T": 19968,
"puct/scale_last": 0.1190591760561926,
"puct/buffer_value/mean": -0.3831989754261772,
"puct/buffer_value/std": 0.015201716201793159,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809408239438133,
"puct/buffer_timestep/mean": 18.6984126984127,
"puct/buffer_timestep/std": 11.387872859938527,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 38.0,
"puct/buffer_construction_len/mean": 79.28095238095239,
"puct/buffer_construction_len/std": 8.844087921117847,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.3809411172476924,
"puct/sampled_value/std": 2.3808762008556907e-07,
"puct/sampled_value/min": -0.38094142084117394,
"puct/sampled_value/max": -0.38094084539764167,
"puct/sampled_timestep/mean": 38.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 38.0,
"puct/sampled_timestep/max": 38.0,
"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": 3696.977634191513,
"env/all/ac_tokens_per_turn": 9589.29296875,
"env/all/ob_tokens_per_turn": 1284.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4909718,
"env/all/total_ob_tokens": 657408,
"env/all/time/sampling_mean": 302.92697605956346,
"env/all/time/sampling_max": 439.1481738090515,
"env/all/time/env_step_mean": 1433.2168450229801,
"env/all/time/env_step_max": 3294.4109456539154,
"env/all/reward/mean": 0.6004366124288499,
"env/all/reward/max": 2.6250794247997704,
"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": 0.6004366124288499,
"env/all/correctness": 0.23046875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3846307195502639,
"env/all/raw_score/min": 0.3809408447995746,
"env/all/raw_score/max": 0.4999999657070216,
"env/all/initial_raw_score": -0.38094111724769253,
"env/all/initial_raw_score/min": -0.38094142084117394,
"env/all/initial_raw_score/max": -0.38094084539764167,
"env/all/msg": "Success; raw_score=0.3809408450624798",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\nfrom math import inf\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n max_corr = np.max(corr[n_points - 1 : 2 * n_points - 1])\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n import random\n random.seed(seed)\n\n # Use the provided initial_h_values if available\n if 'initial_h_values' in globals() and len(initial_h_values) > 0:\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2\n initial_guess = initial_h_values.copy()\n else:\n n_points = 80\n dx = 2.0 / n_points\n required_sum = n_points / 2\n # Generate an alternating initial guess: 1s and 0s\n initial_guess = np.array([1.0 if i % 2 == 0 else 0.0 for i in range(n_points)])\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(x):\n h = np.array(x)\n sum_h = np.sum(h)\n # Penalize constraint violation with a reduced penalty coefficient\n penalty = 1e4 * (sum_h - required_sum) ** 2\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Optimized differential evolution setup\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin', # Better convergence for structured spaces\n popsize=60, # Moderate population for balance\n maxiter=1500, # Sufficient iterations for refinement\n tol=1e-6,\n mutation=(0.2, 0.5), # Narrower mutation for fine-tuning\n recombination=0.9, # Higher recombination for diversity\n seed=seed,\n x0=initial_guess\n )\n\n best_h = result.x\n best_c5 = evaluate_c5(best_h, dx)\n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 302.92697605956346,
"env/all/time/policy/min": 146.0624589920044,
"env/all/time/policy/max": 439.1481738090515,
"env/all/time/env_step": 1433.2168450229801,
"env/all/time/env_step/min": 0.0057222843170166016,
"env/all/time/env_step/max": 3294.4109456539154,
"env/all/time/reward_compute": 3.580935299396515e-07,
"env/all/time/reward_compute/min": 2.7194619178771973e-07,
"env/all/time/reward_compute/max": 5.885958671569824e-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": 0.03318026661872864,
"advantage/min": -0.9873626232147217,
"advantage/max": 6.439675331115723,
"time/assemble_training_data": 8.775553464889526,
"time/kl_vs_base": 79.62511539459229,
"kl_policy_base": 0.0008816916379146278,
"time/train": 551.5873346328735,
"time/save_checkpoint": 15.13869595527649,
"time/total": 4353.662322998047
}
[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
[2026-07-10T08:20:19+00:00] job=1813610 node=node-29 ngpu=6 ntrain=2 replicas=4 flash_attn=yes