Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac2-s2-step25
ttt-discover-v3-qwen3-8b-ac2-s2-step25 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 ac2. Checkpoint saved after training step 24 (0-indexed). Strict upstream eval parity: 1100s hard…
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 ac2. Checkpoint saved
after training step 24 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 24,
"progress/batch": 24,
"optim/lr": 4e-05,
"progress/done_frac": 0.5,
"puct/buffer_size": 389,
"puct/sampled_size": 8,
"puct/T": 12288,
"puct/scale_last": 0.280808686375848,
"puct/buffer_value/mean": 0.9305355322913673,
"puct/buffer_value/std": 0.04571862367740124,
"puct/buffer_value/min": 0.6666666666666636,
"puct/buffer_value/max": 0.9474753530425146,
"puct/buffer_timestep/mean": 11.308483290488432,
"puct/buffer_timestep/std": 7.064335641763451,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 23.0,
"puct/buffer_construction_len/mean": 3737.5192802056554,
"puct/buffer_construction_len/std": 2042.9116369948915,
"puct/buffer_construction_len/min": 1024.0,
"puct/buffer_construction_len/max": 32768.0,
"puct/sampled_value/mean": 0.947475282082322,
"puct/sampled_value/std": 4.695494523289946e-08,
"puct/sampled_value/min": 0.9474751907460986,
"puct/sampled_value/max": 0.9474753530425146,
"puct/sampled_timestep/mean": 23.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 23.0,
"puct/sampled_timestep/max": 23.0,
"puct/sampled_construction_len/mean": 4096.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 4096.0,
"puct/sampled_construction_len/max": 4096.0,
"time/sampling": 4331.286442518234,
"env/all/ac_tokens_per_turn": 8684.11328125,
"env/all/ob_tokens_per_turn": 3641.875,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4446266,
"env/all/total_ob_tokens": 1864640,
"env/all/time/sampling_mean": 326.6318920464255,
"env/all/time/sampling_max": 512.2863137722015,
"env/all/time/env_step_mean": 1753.0260699004866,
"env/all/time/env_step_max": 3895.7429161071777,
"env/all/reward/mean": 0.6195211723830559,
"env/all/reward/max": 0.947505900323792,
"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": 0.6195211723830559,
"env/all/correctness": 0.685546875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.9036890035900988,
"env/all/raw_score/min": 0.02348644064112007,
"env/all/raw_score/max": 0.947505900323792,
"env/all/initial_raw_score": 0.947475282082322,
"env/all/initial_raw_score/min": 0.9474751907460986,
"env/all/initial_raw_score/max": 0.9474753530425146,
"env/all/msg": "Success; raw_score=0.9450322387093714",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom typing import Tuple, List\nimport random\nimport time\nfrom scipy.optimize import minimize\n\ndef _simpson_l2sq(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n m = conv.size\n if m == 0:\n return 0.0, np.zeros_like(conv)\n dx = 1.0 / (m + 1)\n y = np.zeros(m + 2, dtype=conv.dtype)\n y[0] = 0.0\n y[1:-1] = conv\n y[-1] = 0.0\n lhs = y[:-1]\n rhs = y[1:]\n l2_sq = (dx / 3.0) * np.sum(lhs * lhs + lhs * rhs + rhs * rhs)\n grad_y = (dx / 3.0) * (4.0 * y + np.roll(y, 1) + np.roll(y, -1))\n grad_conv = grad_y[1:-1]\n return float(l2_sq), grad_conv\n\ndef _l1(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n m = conv.size\n dx = 1.0 / (m + 1) if m > 0 else 1.0\n val = dx * float(np.sum(conv)) if m > 0 else 0.0\n grad = np.full_like(conv, dx)\n return val, grad\n\ndef _linf(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n if conv.size == 0:\n return 0.0, np.zeros_like(conv)\n m = float(np.max(conv))\n if m <= 0.0:\n return m, np.zeros_like(conv)\n mask = conv == m\n count = int(mask.sum())\n if count == 0:\n return m, np.zeros_like(conv)\n grad = mask.astype(conv.dtype) / count\n return m, grad\n\ndef _objective_and_grad_conv(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n l2_sq, g_l2 = _simpson_l2sq(conv)\n l1, g_l1 = _l1(conv)\n linf, g_linf = _linf(conv)\n if l1 <= 0.0 or linf <= 0.0:\n return 0.0, np.zeros_like(conv)\n denom = l1 * linf\n c_value = l2_sq / denom\n num_grad = g_l2 * denom - l2_sq * (g_l1 * linf + l1 * g_linf)\n g_conv = num_grad / (denom * denom)\n return float(c_value), g_conv\n\ndef _grad_h_from_conv_grad(h: np.ndarray, g_conv: np.ndarray) -> np.ndarray:\n h_rev = h[::-1]\n g_h = np.convolve(g_conv, h_rev, mode=\"valid\")\n return 2.0 * g_h\n\ndef _upsample_1d(h: np.ndarray) -> np.ndarray:\n n = h.shape[0]\n x_old = np.linspace(-0.5, 0.5, n)\n x_new = np.linspace(-0.5, 0.5, 2 * n)\n return np.interp(x_new, x_old, h)\n\ndef construct_function():\n \"\"\"\n Optimizes a sequence of non-negative heights using a refined multi-scale approach with a triangular initial guess,\n adaptive noise and learning rates, and a more efficient final refinement step to improve the ratio of C_lower_bound.\n \"\"\"\n np.random.seed(42)\n \n height_sequence_1 = globals().get(\"height_sequence_1\", None)\n target_length = 4096\n initial_n = 64\n h = np.zeros(target_length, dtype=np.float32)\n\n if height_sequence_1 is not None:\n initial_h = np.array(height_sequence_1, dtype=np.float32)\n initial_n = min(len(initial_h), target_length)\n x_old = np.linspace(-0.5, 0.5, len(initial_h))\n x_new = np.linspace(-0.5, 0.5, initial_n)\n h = np.interp(x_new, x_old, initial_h)\n else:\n x = np.linspace(-0.5, 0.5, target_length)\n triangle = 1.0 - 2.0 * np.abs(x)\n triangle = np.clip(triangle, 0.0, 1000.0)\n x_old = np.linspace(-0.5, 0.5, target_length)\n x_new = np.linspace(-0.5, 0.5, initial_n)\n h = np.interp(x_new, x_old, triangle)\n h = np.pad(h, (0, target_length - initial_n), 'constant', constant_values=0)\n\n learning_rate_initial = 0.05\n noise_scale_initial = 300.0\n noise_decay = 0.999\n learning_rate_decay = 0.998\n steps_per_scale = 500000\n upscale_steps = 3\n\n start_time = time.time()\n\n current_length = initial_n\n scale_factor = 2\n\n for _ in range(upscale_steps):\n if current_length < target_length:\n h = _upsample_1d(h)\n current_length *= 2\n h_sum = np.sum(h)\n if h_sum < 0.01:\n h = np.clip(h + (0.01 - h_sum) / h.shape[0], 0.0, 1000.0)\n h_opt = np.copy(h)\n for step in range(steps_per_scale):\n clipped_h = np.clip(h_opt, 0.0, None)\n conv = np.convolve(clipped_h, clipped_h, mode='full')\n obj_val, grad_conv = _objective_and_grad_conv(conv)\n grad_h = _grad_h_from_conv_grad(clipped_h, grad_conv)\n noise_scale = noise_scale_initial * (noise_decay ** step)\n learning_rate_current = learning_rate_initial * (learning_rate_decay ** step)\n noise = noise_scale * np.random.normal(size=h_opt.shape)\n h_opt = np.clip(h_opt + learning_rate_current * grad_h + noise, 0.0, 1000.0)\n h_sum = np.sum(h_opt)\n if h_sum < 0.01:\n h_opt = np.clip(h_opt + (0.01 - h_sum) / h_opt.shape[0], 0.0, 1000.0)\n remaining_time = 1000 - (time.time() - start_time)\n if remaining_time < 10:\n print(f\"Time remaining: {remaining_time} seconds. Performing final refinement.\")\n break\n h = h_opt\n\n h_initial = np.copy(h)\n learning_rate = 0.1\n beta1 = 0.9\n beta2 = 0.999\n epsilon = 1e-8\n m = np.zeros_like(h)\n v = np.zeros_like(h)\n t = 0\n num_steps = 100000\n for step in range(num_steps):\n clipped_h = np.clip(h, 0.0, None)\n conv = np.convolve(clipped_h, clipped_h, mode='full')\n obj_val, grad_conv = _objective_and_grad_conv(conv)\n grad_h = _grad_h_from_conv_grad(clipped_h, grad_conv)\n t += 1\n grad = grad_h\n m = beta1 * m + (1 - beta1) * grad\n v = beta2 * v + (1 - beta2) * grad ** 2\n m_hat = m / (1 - beta1 ** t)\n v_hat = v / (1 - beta2 ** t)\n learning_rate_current = learning_rate * (1 / (1 + step))\n h = np.clip(h + learning_rate_current * m_hat / (np.sqrt(v_hat) + epsilon), 0.0, 1000.0)\n h_sum = np.sum(h)\n if h_sum < 0.01:\n h = np.clip(h + (0.01 - h_sum) / h.shape[0], 0.0, 1000.0)\n remaining_time = 1000 - (time.time() - start_time)\n if remaining_time < 10:\n print(f\"Time remaining: {remaining_time} seconds. Ending final refinement.\")\n break\n\n h_final = np.clip(h, 0.0, 1000.0)\n heights = h_final.tolist()\n r_value = evaluate_sequence(heights)\n print(f\"Final C2 lower bound: {r_value}\")\n return heights\n```",
"env/all/time/policy": 326.6318920464255,
"env/all/time/policy/min": 87.059903383255,
"env/all/time/policy/max": 512.2863137722015,
"env/all/time/env_step": 1753.0260699004866,
"env/all/time/env_step/min": 0.006566762924194336,
"env/all/time/env_step/max": 3895.7429161071777,
"env/all/time/reward_compute": 2.6496127247810364e-07,
"env/all/time/reward_compute/min": 2.086162567138672e-07,
"env/all/time/reward_compute/max": 4.3958425521850586e-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": 0.012045965529978275,
"advantage/min": -1.0,
"advantage/max": 1.2549464702606201,
"time/assemble_training_data": 6.972659587860107,
"time/kl_vs_base": 97.37902474403381,
"kl_policy_base": 0.000694518443197012,
"time/train": 595.759051322937,
"time/save_checkpoint": 11.589547157287598,
"time/total": 5047.493044376373
}
[2026-07-09T06:42:12+00:00] job=1812634 node=node-30 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:26:33+00:00] job=1812736 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:27:32+00:00] job=1813133 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T16:26:04+00:00] job=1813134 node=node-4 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-11T01:56:04+00:00] job=1821290 node=node-4 ngpu=3 ntrain=1 replicas=2 flash_attn=yes