Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step43
ttt-discover-v3-qwen3-8b-erdos-step43 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 42 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 11, 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 42 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 42,
"progress/batch": 42,
"optim/lr": 4e-05,
"progress/done_frac": 0.86,
"puct/buffer_size": 676,
"puct/sampled_size": 8,
"puct/T": 21504,
"puct/scale_last": 0.1190591760561926,
"puct/buffer_value/mean": -0.3831971337433007,
"puct/buffer_value/std": 0.015185668052820235,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809408239438133,
"puct/buffer_timestep/mean": 20.144970414201183,
"puct/buffer_timestep/std": 12.229547163149286,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 41.0,
"puct/buffer_construction_len/mean": 79.32988165680473,
"puct/buffer_construction_len/std": 8.53979861243955,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.3809408423608599,
"puct/sampled_value/std": 2.2013907330978643e-09,
"puct/sampled_value/min": -0.3809408448742484,
"puct/sampled_value/max": -0.380940837884394,
"puct/sampled_timestep/mean": 41.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 41.0,
"puct/sampled_timestep/max": 41.0,
"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": 3618.2902040481567,
"env/all/ac_tokens_per_turn": 9571.666015625,
"env/all/ob_tokens_per_turn": 1298.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4900693,
"env/all/total_ob_tokens": 664640,
"env/all/time/sampling_mean": 299.4260990046896,
"env/all/time/sampling_max": 426.0151116847992,
"env/all/time/env_step_mean": 1321.6914713927545,
"env/all/time/env_step_max": 3220.8959126472473,
"env/all/reward/mean": 0.8207998857422187,
"env/all/reward/max": 2.6250798248676035,
"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": 0.8207998857422187,
"env/all/correctness": 0.314453125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3835815461845279,
"env/all/raw_score/min": 0.3809407867433658,
"env/all/raw_score/max": 0.5,
"env/all/initial_raw_score": -0.3809408423608599,
"env/all/initial_raw_score/min": -0.3809408448742484,
"env/all/initial_raw_score/max": -0.380940837884394,
"env/all/msg": "Success; raw_score=0.38094167588615574",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\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 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 provided initial guess if it exists\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 initial_guess = np.zeros(n_points)\n initial_guess[:n_points // 2] = 1.0 # 0.5 density\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(x):\n h = np.array(x)\n max_c5 = evaluate_c5(h, dx)\n sum_h = np.sum(h)\n penalty = 1e2 * (sum_h - required_sum) ** 2 # Reduced penalty\n return max_c5 + penalty\n\n # Run differential evolution with refined parameters\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=60,\n maxiter=1000,\n tol=1e-6,\n mutation=(0.2, 0.5),\n recombination=0.9,\n seed=seed,\n x0=initial_guess\n )\n\n best_h_de = result_de.x\n best_c5_de = evaluate_c5(best_h_de, dx)\n\n # Local refinement with L-BFGS-B\n def objective_local(x):\n h = np.array(x)\n max_c5 = evaluate_c5(h, dx)\n sum_h = np.sum(h)\n penalty = 1e2 * (sum_h - required_sum) ** 2\n return max_c5 + penalty\n\n result_local = minimize(\n objective_local,\n best_h_de,\n method='L-BFGS-B',\n bounds=bounds,\n tol=1e-6\n )\n\n best_h = result_local.x\n best_c5 = evaluate_c5(best_h, dx)\n\n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 299.4260990046896,
"env/all/time/policy/min": 108.70718836784363,
"env/all/time/policy/max": 426.0151116847992,
"env/all/time/env_step": 1321.6914713927545,
"env/all/time/env_step/min": 0.006517887115478516,
"env/all/time/env_step/max": 3220.8959126472473,
"env/all/time/reward_compute": 3.082677721977234e-07,
"env/all/time/reward_compute/min": 1.9744038581848145e-07,
"env/all/time/reward_compute/max": 4.470348358154297e-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": 0.03299068287014961,
"advantage/min": -1.0,
"advantage/max": 14.11259937286377,
"time/assemble_training_data": 6.80657172203064,
"time/kl_vs_base": 81.43360495567322,
"kl_policy_base": 0.0009334296337328851,
"time/train": 550.3178515434265,
"time/save_checkpoint": 15.187975406646729,
"time/total": 4274.142314910889
}
[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