Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step45
ttt-discover-v3-qwen3-8b-erdos-step45 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 44 (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 44 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 44,
"progress/batch": 44,
"optim/lr": 4e-05,
"progress/done_frac": 0.9,
"puct/buffer_size": 708,
"puct/sampled_size": 8,
"puct/T": 22528,
"puct/scale_last": 0.1190592132566401,
"puct/buffer_value/mean": -0.3831748211783583,
"puct/buffer_value/std": 0.014984800480319135,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809407867433658,
"puct/buffer_timestep/mean": 21.15536723163842,
"puct/buffer_timestep/std": 12.8210692153811,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 43.0,
"puct/buffer_construction_len/mean": 79.38841807909604,
"puct/buffer_construction_len/std": 8.381627155195625,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38094083480391494,
"puct/sampled_value/std": 5.4329689622127954e-09,
"puct/sampled_value/min": -0.3809408412803688,
"puct/sampled_value/max": -0.38094082442404065,
"puct/sampled_timestep/mean": 43.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 43.0,
"puct/sampled_timestep/max": 43.0,
"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": 3937.5610086917877,
"env/all/ac_tokens_per_turn": 9294.88671875,
"env/all/ob_tokens_per_turn": 1260.25,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4758982,
"env/all/total_ob_tokens": 645248,
"env/all/time/sampling_mean": 287.0115217366256,
"env/all/time/sampling_max": 401.5274567604065,
"env/all/time/env_step_mean": 1512.7813044744544,
"env/all/time/env_step_max": 3532.294144630432,
"env/all/reward/mean": 0.8118315255593225,
"env/all/reward/max": 2.625079527117616,
"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": 0.8118315255593225,
"env/all/correctness": 0.310546875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3829342534311112,
"env/all/raw_score/min": 0.38094082995162526,
"env/all/raw_score/max": 0.5,
"env/all/initial_raw_score": -0.38094083480391494,
"env/all/initial_raw_score/min": -0.3809408412803688,
"env/all/initial_raw_score/max": -0.38094082442404065,
"env/all/msg": "Evaluation timed out after 1100 seconds.",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\nimport time\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n one_minus_h = 1.0 - h\n # Compute full correlation (includes overlap regions)\n corr = np.correlate(h, one_minus_h, mode='full')\n # Focus on main overlap region (avoid edge effects)\n # Overlap around k=0.5, 1.0, etc. is where the max is likely\n # We take the middle 100 points for n_points=128 as most representative\n max_corr = np.max(corr[n_points - 50 : n_points + 50])\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 n_points = 128\n dx = 2.0 / n_points\n required_sum = n_points / 2\n initial_guess = np.zeros(n_points)\n initial_guess[:32] = 1.0\n initial_guess[32:96] = 0.5\n initial_guess[96:] = 0.0\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 penalty = 100.0 * (sum_h - required_sum) ** 2 # Reduced penalty\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Differential Evolution for global optimization\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=100,\n maxiter=750,\n tol=1e-5,\n mutation=(0.4, 0.5),\n recombination=0.8,\n seed=seed,\n x0=initial_guess\n )\n\n best_h = result.x\n best_c5 = evaluate_c5(best_h, dx)\n\n # Local optimization with L-BFGS-B\n def objective_local(h_arr):\n return evaluate_c5(h_arr, dx)\n\n local_result = minimize(\n objective_local,\n best_h,\n method='L-BFGS-B',\n bounds=bounds,\n options={'maxiter': 200, 'ftol': 1e-8}\n )\n\n final_h = local_result.x\n final_c5 = evaluate_c5(final_h, dx)\n\n return final_h, final_c5, n_points\n```",
"env/all/time/policy": 287.0115217366256,
"env/all/time/policy/min": 114.62326550483704,
"env/all/time/policy/max": 401.5274567604065,
"env/all/time/env_step": 1512.7813044744544,
"env/all/time/env_step/min": 0.005221843719482422,
"env/all/time/env_step/max": 3532.294144630432,
"env/all/time/reward_compute": 3.105960786342621e-07,
"env/all/time/reward_compute/min": 2.086162567138672e-07,
"env/all/time/reward_compute/max": 4.76837158203125e-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": 0.026326587423682213,
"advantage/min": -1.0,
"advantage/max": 4.357759952545166,
"time/assemble_training_data": 7.380000829696655,
"time/kl_vs_base": 79.44050621986389,
"kl_policy_base": 0.0009681851370260119,
"time/train": 531.3877754211426,
"time/save_checkpoint": 7.770035982131958,
"time/total": 4564.991193056107
}
[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