Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step20
ttt-discover-v3-qwen3-8b-erdos-step20 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 19 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 16, 2026
Repo size
355 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 19 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 19,
"progress/batch": 19,
"optim/lr": 4e-05,
"progress/done_frac": 0.4,
"puct/buffer_size": 312,
"puct/sampled_size": 8,
"puct/T": 9728,
"puct/scale_last": 0.1190584434086483,
"puct/buffer_value/mean": -0.38539883617328863,
"puct/buffer_value/std": 0.021337147488676628,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809415565913576,
"puct/buffer_timestep/mean": 8.743589743589743,
"puct/buffer_timestep/std": 5.632861224571347,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 18.0,
"puct/buffer_construction_len/mean": 78.54807692307692,
"puct/buffer_construction_len/std": 12.525010728046208,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.3809415570593055,
"puct/sampled_value/std": 9.360742085227836e-10,
"puct/sampled_value/min": -0.38094155953061776,
"puct/sampled_value/max": -0.3809415565913576,
"puct/sampled_timestep/mean": 18.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 18.0,
"puct/sampled_timestep/max": 18.0,
"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": 2099.183831214905,
"env/all/ac_tokens_per_turn": 9674.544921875,
"env/all/ob_tokens_per_turn": 1397.75,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4953367,
"env/all/total_ob_tokens": 715648,
"env/all/time/sampling_mean": 534.8097359854728,
"env/all/time/sampling_max": 766.9931488037109,
"env/all/time/env_step_mean": 353.347635735292,
"env/all/time/env_step_max": 1410.0371947288513,
"env/all/reward/mean": 0.9818798560448678,
"env/all/reward/max": 2.6250745253158194,
"env/all/reward/min": 0.0,
"env/all/format": 0.998046875,
"env/all/format/min": 0.0,
"env/all/format/max": 1.0,
"env/all/reward": 0.9818798560448678,
"env/all/correctness": 0.392578125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.4092552465738802,
"env/all/raw_score/min": 0.38094155579409544,
"env/all/raw_score/max": 1.0,
"env/all/initial_raw_score": -0.3809415570593055,
"env/all/initial_raw_score/min": -0.38094155953061776,
"env/all/initial_raw_score/max": -0.3809415565913576,
"env/all/msg": "Success; raw_score=0.3809415571441932",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\nfrom scipy.fft import fft, ifft\n\ndef evaluate_C5_fft(h, dx):\n h = np.asarray(h)\n h1 = 1.0 - h\n n = len(h)\n len_fft = 2 * n - 1\n h_padded = np.zeros(len_fft, dtype=h.dtype)\n h_padded[:n] = h\n h1_padded = np.zeros(len_fft, dtype=h1.dtype)\n h1_padded[:n] = h1\n fft_h = fft(h_padded)\n fft_h1 = fft(h1_padded)\n corr_fft = fft_h * np.conj(fft_h1)\n corr = np.real(ifft(corr_fft))\n max_corr = np.max(corr)\n return max_corr * dx\n\ndef generate_binary_initial(n_points):\n indices = np.random.choice(n_points, size=n_points // 2, replace=False)\n h = np.zeros(n_points)\n h[indices] = 1\n return h\n\ndef generate_alternating_initial(n_points):\n h = np.zeros(n_points)\n for i in range(n_points):\n if i % 2 == 0:\n h[i] = 1\n else:\n h[i] = 0\n return h\n\ndef generate_uniform_initial(n_points):\n indices = np.sort(np.unique(np.round(np.linspace(0, n_points - 1, n_points // 2))).astype(int))\n h = np.zeros(n_points)\n h[indices] = 1\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 n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points // 2\n initial_guesses = []\n \n # Add multiple initial guess strategies\n initial_guesses.append(initial_h_values.copy())\n initial_guesses.append(generate_binary_initial(n_points))\n initial_guesses.append(generate_alternating_initial(n_points))\n initial_guesses.append(generate_uniform_initial(n_points))\n\n best_h_scaled = None\n best_c5 = float('inf')\n\n # Run DE on each initial guess\n for initial_guess in initial_guesses:\n if np.sum(initial_guess) == 0:\n continue\n # Ensure the initial guess has enough 1s\n if np.sum(initial_guess) < 1:\n continue\n\n # Scale initial guess to satisfy the constraint\n scaled_initial = initial_guess * (required_sum / np.sum(initial_guess))\n\n def objective(x):\n sum_x = np.sum(x)\n if sum_x == 0:\n return float('inf')\n scaled_x = x * (required_sum / sum_x)\n return evaluate_C5_fft(scaled_x, dx)\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Run differential evolution with optimized parameters\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=50,\n maxiter=300,\n tol=1e-6,\n mutation=(0.5, 1.0),\n recombination=0.9,\n disp=False,\n polish=True,\n x0=scaled_initial\n )\n\n # Track best solution\n if result.fun < best_c5:\n best_c5 = result.fun\n best_h_scaled = result.x * (required_sum / np.sum(result.x))\n\n # Local refinement using SLSQP with constraint\n def objective_refine(x):\n sum_x = np.sum(x)\n if sum_x == 0:\n return float('inf')\n return evaluate_C5_fft(x, dx)\n\n bounds_refine = [(0.0, 1.0) for _ in range(n_points)]\n\n # Add constraint: exactly required_sum 1's\n constraints = ({'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum})\n\n result_refine = minimize(\n objective_refine,\n best_h_scaled,\n method='SLSQP',\n bounds=bounds_refine,\n constraints=constraints,\n options={'maxiter': 100, 'disp': False}\n )\n\n best_h_scaled = result_refine.x\n best_c5 = result_refine.fun\n\n return best_h_scaled, best_c5, n_points\n```",
"env/all/time/policy": 534.8097359854728,
"env/all/time/policy/min": 256.3743920326233,
"env/all/time/policy/max": 766.9931488037109,
"env/all/time/env_step": 353.347635735292,
"env/all/time/env_step/min": 0.0051937103271484375,
"env/all/time/env_step/max": 1410.0371947288513,
"env/all/time/reward_compute": 3.264285624027252e-07,
"env/all/time/reward_compute/min": 2.4586915969848633e-07,
"env/all/time/reward_compute/max": 6.48200511932373e-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": 0.0229030828922987,
"advantage/min": -1.0,
"advantage/max": 3.1876096725463867,
"time/assemble_training_data": 8.589319944381714,
"time/kl_vs_base": 128.1708629131317,
"kl_policy_base": 0.0008764734375290573,
"time/train": 1036.802077293396,
"time/save_checkpoint": 11.274418592453003,
"time/total": 3285.1867229938507
}
[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