Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac2-s2-step19
ttt-discover-v3-qwen3-8b-ac2-s2-step19 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 18 (0-indexed). Strict upstream eval parity: 1100s hard…
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 ac2. Checkpoint saved
after training step 18 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 18,
"progress/batch": 18,
"optim/lr": 4e-05,
"progress/done_frac": 0.38,
"puct/buffer_size": 293,
"puct/sampled_size": 8,
"puct/T": 9216,
"puct/scale_last": 0.1910110676940776,
"puct/buffer_value/mean": 0.9259520747682203,
"puct/buffer_value/std": 0.04923036008024519,
"puct/buffer_value/min": 0.6666666666666636,
"puct/buffer_value/max": 0.9474669279340687,
"puct/buffer_timestep/mean": 8.296928327645052,
"puct/buffer_timestep/std": 5.34322550752021,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 17.0,
"puct/buffer_construction_len/mean": 3606.085324232082,
"puct/buffer_construction_len/std": 2326.8536042887285,
"puct/buffer_construction_len/min": 1024.0,
"puct/buffer_construction_len/max": 32768.0,
"puct/sampled_value/mean": 0.9474207891692258,
"puct/sampled_value/std": 3.649366856365714e-05,
"puct/sampled_value/min": 0.9473477964088227,
"puct/sampled_value/max": 0.9474669279340687,
"puct/sampled_timestep/mean": 17.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 17.0,
"puct/sampled_timestep/max": 17.0,
"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": 4550.421635389328,
"env/all/ac_tokens_per_turn": 8601.05078125,
"env/all/ob_tokens_per_turn": 3645.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4403738,
"env/all/total_ob_tokens": 1866560,
"env/all/time/sampling_mean": 325.8045159704052,
"env/all/time/sampling_max": 418.6484844684601,
"env/all/time/env_step_mean": 1877.9345220732503,
"env/all/time/env_step_max": 4138.958982229233,
"env/all/reward/mean": 0.5031069845969315,
"env/all/reward/max": 0.9474733645012318,
"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": 0.5031069845969315,
"env/all/correctness": 0.59765625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.8417999219399638,
"env/all/raw_score/min": 0.023451023384054574,
"env/all/raw_score/max": 0.9474733645012318,
"env/all/initial_raw_score": 0.9474207891692259,
"env/all/initial_raw_score/min": 0.9473477964088227,
"env/all/initial_raw_score/max": 0.9474669279340687,
"env/all/msg": "Success; raw_score=0.9474625815493276",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom typing import Tuple, List\nimport random\nimport time\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 _single_candidate_finetune(h0: np.ndarray, lr=5e-5, steps=200000) -> Tuple[np.ndarray, float]:\n h = h0.astype(np.float32).copy()\n opt = _Adam(h.shape, lr=lr, dtype=h.dtype)\n last_c = 0.0\n for _ in range(steps):\n h_clip = np.clip(h, 0.0, None)\n conv = np.convolve(h_clip, h_clip, mode=\"full\")\n c_val, g_conv = _objective_and_grad_conv(conv)\n g_h = _grad_h_from_conv_grad(h_clip, g_conv)\n h = np.clip(opt.step(h, g_h.astype(h.dtype)), 0.0, None)\n last_c = c_val\n return h, float(last_c)\n\nclass _Adam:\n def __init__(self, shape, lr=3e-2, beta1=0.9, beta2=0.999, eps=1e-8, dtype=np.float32):\n self.m = np.zeros(shape, dtype=dtype)\n self.v = np.zeros(shape, dtype=dtype)\n self.t = 0\n self.lr = lr\n self.b1 = beta1\n self.b2 = beta2\n self.eps = eps\n\n def step(self, params, grad):\n self.t += 1\n self.m = self.b1 * self.m + (1 - self.b1) * grad\n self.v = self.b2 * self.v + (1 - self.b2) * (grad * grad)\n m_hat = self.m / (1 - self.b1 ** self.t)\n v_hat = self.v / (1 - self.b2 ** self.t)\n return params + self.lr * m_hat / (np.sqrt(v_hat) + self.eps)\n\ndef construct_function():\n \"\"\"\n Optimizes a sequence of non-negative heights using a combination of multi-scale\n optimization with adaptive noise and learning rate strategies. The initial sequence\n is designed as a Gaussian to better balance L2, L1, and L\u221e norms. The algorithm \n dynamically adjusts noise and learning rates to explore the search space efficiently.\n \"\"\"\n np.random.seed(42)\n \n # Load previous best sequence\n height_sequence_1 = globals().get(\"height_sequence_1\", None)\n target_length = 4096\n initial_n = 64 # Starting sequence length\n h = np.zeros(target_length, dtype=np.float32)\n\n if height_sequence_1 is not None:\n # Use the previous best and upscale\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 # Generate a Gaussian-shaped initial sequence\n h = np.exp(-16 * (np.linspace(-0.5, 0.5, target_length) ** 2))\n h = np.clip(h, 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) / target_length, 0.0, 1000.0)\n # Trim to initial_n (64) and pad\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, h)\n h = np.pad(h, (0, target_length - initial_n), 'constant', constant_values=0)\n \n # Parameters\n learning_rate = 1.0\n noise_scale_initial = 50.0 # Larger initial noise\n noise_decay = 0.9995 # Slightly slower decay\n learning_rate_decay = 0.9995 # Slower learning rate decay\n max_steps = 1_000_000\n upscale_steps = 500_000\n refine_steps = 300_000\n\n start_time = time.time()\n\n # Multi-scale optimization with extended upscale and adaptive training\n current_length = initial_n\n scale_factor = 2\n\n for _ in range(6): # Upscale 6 times to reach 4096\n if current_length < target_length:\n h = _upsample_1d(h)\n current_length *= 2\n # Ensure sum is above 0.01\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 \n # Initial optimization with noise\n h_opt = np.copy(h)\n for step in range(100000): # Reduced steps per upscale\n # Clip to non-negative\n clipped_h = np.clip(h_opt, 0.0, None)\n \n # Compute convolution\n conv = np.convolve(clipped_h, clipped_h, mode='full')\n \n # Compute objective and gradient of conv\n obj_val, grad_conv = _objective_and_grad_conv(conv)\n \n # Compute gradient with respect to h\n grad_h = _grad_h_from_conv_grad(clipped_h, grad_conv)\n \n # Adjust noise and learning rate based on step\n noise_scale = noise_scale_initial * (noise_decay ** step)\n learning_rate_current = learning_rate * (learning_rate_decay ** step)\n \n # Add noise\n noise = noise_scale * np.random.normal(size=h_opt.shape)\n \n # Update h\n h_opt = np.clip(h_opt + learning_rate_current * grad_h + noise, 0.0, 1000.0)\n \n # Ensure sum is above 0.01\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 \n # Check time remaining\n remaining_time = 1000 - (time.time() - start_time)\n if remaining_time < 5:\n print(f\"Time remaining: {remaining_time} seconds. Performing final refinement.\")\n break\n # Assign optimized h\n h = h_opt\n \n # Final refined optimization with smaller learning rate\n h_refined, _ = _single_candidate_finetune(h, lr=1e-4, steps=refine_steps)\n \n # Final check\n h_final = np.clip(h_refined, 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": 325.8045159704052,
"env/all/time/policy/min": 171.42873811721802,
"env/all/time/policy/max": 418.6484844684601,
"env/all/time/env_step": 1877.9345220732503,
"env/all/time/env_step/min": 0.006897449493408203,
"env/all/time/env_step/max": 4138.958982229233,
"env/all/time/reward_compute": 2.6868656277656555e-07,
"env/all/time/reward_compute/min": 2.086162567138672e-07,
"env/all/time/reward_compute/max": 3.986060619354248e-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": 0.02164783701300621,
"advantage/min": -1.0,
"advantage/max": 2.954786777496338,
"time/assemble_training_data": 10.634697914123535,
"time/kl_vs_base": 102.2416763305664,
"kl_policy_base": 0.0006905561313033104,
"time/train": 594.212641954422,
"time/save_checkpoint": 17.552107572555542,
"time/total": 5278.721784353256
}
[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