Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step32
ttt-discover-v3-qwen3-8b-erdos-step32 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 31 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 10, 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 31 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 31,
"progress/batch": 31,
"optim/lr": 4e-05,
"progress/done_frac": 0.64,
"puct/buffer_size": 503,
"puct/sampled_size": 8,
"puct/T": 15872,
"puct/scale_last": 0.1190591760561926,
"puct/buffer_value/mean": -0.3837689615194966,
"puct/buffer_value/std": 0.016965490681221367,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809408239438133,
"puct/buffer_timestep/mean": 14.715705765407554,
"puct/buffer_timestep/std": 9.079102994339015,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 30.0,
"puct/buffer_construction_len/mean": 79.09940357852882,
"puct/buffer_construction_len/std": 9.889552072816112,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.3809414358104517,
"puct/sampled_value/std": 2.31390653055909e-07,
"puct/sampled_value/min": -0.380941530374009,
"puct/sampled_value/max": -0.3809408239438133,
"puct/sampled_timestep/mean": 30.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 30.0,
"puct/sampled_timestep/max": 30.0,
"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": 4630.548825502396,
"env/all/ac_tokens_per_turn": 9578.521484375,
"env/all/ob_tokens_per_turn": 1279.25,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4904203,
"env/all/total_ob_tokens": 654976,
"env/all/time/sampling_mean": 295.86934563703835,
"env/all/time/sampling_max": 407.67772364616394,
"env/all/time/env_step_mean": 1980.867622022517,
"env/all/time/env_step_max": 4226.720216751099,
"env/all/reward/mean": 0.3468099290501331,
"env/all/reward/max": 2.625074933798669,
"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": 0.3468099290501331,
"env/all/correctness": 0.1328125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.38345162991900816,
"env/all/raw_score/min": 0.38094149651651277,
"env/all/raw_score/max": 0.5,
"env/all/initial_raw_score": -0.3809414358104517,
"env/all/initial_raw_score/min": -0.380941530374009,
"env/all/initial_raw_score/max": -0.3809408239438133,
"env/all/msg": "C5 mismatch: reported 0.22722570, computed 0.66957285",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\nfrom scipy.signal import correlate\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n one_minus_h = 1.0 - h\n corr = correlate(h, one_minus_h, mode='full')\n corr_valid = corr[n_points - 1 : 2 * n_points - 1]\n max_c5 = np.max(corr_valid) * dx\n return max_c5\n\ndef generate_random_initial_guess(n_points):\n h = np.random.rand(n_points)\n required_sum = n_points / 2.0\n current_sum = np.sum(h)\n h = h * (required_sum / current_sum)\n return h\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_h_values if present, else generate a random initial guess\n if initial_h_values is None:\n n_points = 80\n h_initial = generate_random_initial_guess(n_points)\n else:\n n_points = len(initial_h_values)\n h_initial = np.array(initial_h_values)\n # Scale to meet the integral constraint\n required_sum = n_points / 2.0\n current_sum = np.sum(h_initial)\n h_initial = h_initial * (required_sum / current_sum)\n\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\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 # Stronger penalty for constraint violation\n penalty = 1e6 * (sum_h - required_sum) ** 2 # Aggressive constraint enforcement\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Run Differential Evolution with tuned parameters\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='rand1bin', # Encourages diversity\n popsize=250, # Larger population for better exploration\n maxiter=100, # Faster convergence within the budget\n tol=1e-6,\n mutation=(0.5, 0.9), # Balanced mutation range\n recombination=0.8,\n seed=seed,\n x0=h_initial\n )\n\n best_h_de = result_de.x\n best_c5_de = evaluate_c5(best_h_de, dx)\n\n # Refine with SLSQP to improve final result\n def objective_slsqp(x):\n return evaluate_c5(x, dx)\n\n constraints = [{'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum}]\n result_slsqp = minimize(\n objective_slsqp,\n best_h_de,\n bounds=bounds,\n constraints=constraints,\n method='SLSQP',\n tol=1e-8 # Tighter tolerance\n )\n\n best_h_final = result_slsqp.x\n best_c5_final = evaluate_c5(best_h_final, dx)\n\n return best_h_final, best_c5_final, n_points\n```",
"env/all/time/policy": 295.86934563703835,
"env/all/time/policy/min": 148.99914503097534,
"env/all/time/policy/max": 407.67772364616394,
"env/all/time/env_step": 1980.867622022517,
"env/all/time/env_step/min": 0.0048787593841552734,
"env/all/time/env_step/max": 4226.720216751099,
"env/all/time/reward_compute": 2.966262400150299e-07,
"env/all/time/reward_compute/min": 1.4528632164001465e-07,
"env/all/time/reward_compute/max": 4.172325134277344e-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": 0.046730220317840576,
"advantage/min": -0.7496095895767212,
"advantage/max": 14.112607955932617,
"time/assemble_training_data": 7.692726135253906,
"time/kl_vs_base": 79.90849900245667,
"kl_policy_base": 0.0008702584891580045,
"time/train": 550.7006487846375,
"time/save_checkpoint": 15.378533601760864,
"time/total": 5285.671457529068
}
[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