Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac2-s2-step13
ttt-discover-v3-qwen3-8b-ac2-s2-step13 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 12 (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 12 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 12,
"progress/batch": 12,
"optim/lr": 4e-05,
"progress/done_frac": 0.26,
"puct/buffer_size": 197,
"puct/sampled_size": 8,
"puct/T": 6144,
"puct/scale_last": 0.18949366619241537,
"puct/buffer_value/mean": 0.9166524225151236,
"puct/buffer_value/std": 0.05757056815513018,
"puct/buffer_value/min": 0.6666666666666636,
"puct/buffer_value/max": 0.9459495264324065,
"puct/buffer_timestep/mean": 5.274111675126903,
"puct/buffer_timestep/std": 3.6267463635465624,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 11.0,
"puct/buffer_construction_len/mean": 3346.5532994923856,
"puct/buffer_construction_len/std": 2786.181547874715,
"puct/buffer_construction_len/min": 1024.0,
"puct/buffer_construction_len/max": 32768.0,
"puct/sampled_value/mean": 0.9447956392262008,
"puct/sampled_value/std": 0.0007033672038342041,
"puct/sampled_value/min": 0.9439469911638841,
"puct/sampled_value/max": 0.9459495264324065,
"puct/sampled_timestep/mean": 11.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 11.0,
"puct/sampled_timestep/max": 11.0,
"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": 5416.706344842911,
"env/all/ac_tokens_per_turn": 8366.15234375,
"env/all/ob_tokens_per_turn": 3899.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4283470,
"env/all/total_ob_tokens": 1996480,
"env/all/time/sampling_mean": 314.8412881754339,
"env/all/time/sampling_max": 430.1409065723419,
"env/all/time/env_step_mean": 2398.673213653732,
"env/all/time/env_step_max": 5008.578496217728,
"env/all/reward/mean": 0.23994552244448503,
"env/all/reward/max": 0.9457399588945222,
"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": 0.23994552244448503,
"env/all/correctness": 0.271484375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.8838281114501894,
"env/all/raw_score/min": 0.37298536193673815,
"env/all/raw_score/max": 0.9457399588945222,
"env/all/initial_raw_score": 0.9447956392262007,
"env/all/initial_raw_score/min": 0.9439469911638841,
"env/all/initial_raw_score/max": 0.9459495264324065,
"env/all/msg": "Evaluation timed out after 1100 seconds.",
"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-3, steps=500000) -> 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 for maximizing the evaluation score,\n using a combination of diverse initial sequences, adaptive gradient ascent with enhanced exploration,\n and multi-scale refinement with multiple upsampling steps. The algorithm explores multiple structured\n and random initial sequences, and refines them with increased refinement steps and adjusted hyperparameters.\n \"\"\"\n np.random.seed(0)\n \n # Load previous best sequence\n height_sequence_1 = globals().get(\"height_sequence_1\", None)\n initial_n = 1024 # Starting sequence length\n initial_candidates = []\n \n if height_sequence_1 is not None:\n # Use previous best and upscale\n initial_h = np.array(height_sequence_1, dtype=np.float32)\n initial_n = min(len(initial_h), 2048)\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 initial_candidates.append(h)\n else:\n # Generate diverse initial sequences\n # Random\n initial_candidates.append(np.random.uniform(0.1, 1.0, initial_n))\n # Single peak\n single_peak = np.zeros(initial_n, dtype=np.float32)\n single_peak[initial_n // 2] = 1.0\n initial_candidates.append(single_peak)\n # Multi-peak comb filter\n def generate_comb(n, spacing=4):\n seq = np.zeros(n, dtype=np.float32)\n for i in range(0, n, spacing):\n seq[i] = 1.0\n return seq\n initial_candidates.append(generate_comb(initial_n, spacing=8))\n # Multi-peak\n multi_peak = np.zeros(initial_n, dtype=np.float32)\n for i in range(0, initial_n, initial_n // 4):\n multi_peak[i] = 1.0\n initial_candidates.append(multi_peak)\n # Flat\n flat_seq = np.ones(initial_n, dtype=np.float32) * 0.1\n initial_candidates.append(flat_seq)\n # Gaussian-like\n def generate_gaussian(n, peak=1.0, std_dev=0.3):\n x = np.linspace(-0.5, 0.5, n)\n y = np.exp(-((x) ** 2) / (2 * std_dev ** 2)) * peak\n return np.clip(y, 0.0, 1000.0)\n initial_candidates.append(generate_gaussian(initial_n))\n # Exponential decay\n def generate_exponential(n, initial_value=1.0, decay_rate=0.2):\n x = np.linspace(0, n-1, n)\n y = initial_value * np.exp(-decay_rate * x)\n return np.clip(y, 0.0, 1000.0)\n initial_candidates.append(generate_exponential(initial_n))\n # Symmetric\n def generate_symmetric(n):\n half = n // 2\n seq = np.zeros(n, dtype=np.float32)\n seq[:half] = np.random.uniform(0.1, 1.0, half)\n seq[half:] = seq[:half][::-1]\n return seq\n initial_candidates.append(generate_symmetric(initial_n))\n # Double peak\n def generate_double_peak(n):\n seq = np.zeros(n, dtype=np.float32)\n seq[n//4] = 1.0\n seq[3*n//4] = 1.0\n return seq\n initial_candidates.append(generate_double_peak(initial_n))\n # Additional candidates\n def generate_abs_sine(n, freq=2, amp=0.5):\n x = np.linspace(-0.5, 0.5, n)\n y = amp * np.abs(np.sin(freq * np.pi * x))\n return y\n initial_candidates.append(generate_abs_sine(initial_n))\n \n def generate_triangular(n):\n x = np.linspace(-0.5, 0.5, n)\n y = np.abs(x) * 2 * (n - 1) / 1.0 # Normalize to between 0 and 1\n y = np.clip(y, 0.0, 1.0)\n return y\n initial_candidates.append(generate_triangular(initial_n))\n \n def generate_random_peaks(n, num_peaks=5):\n seq = np.zeros(n, dtype=np.float32)\n positions = np.random.choice(range(n), num_peaks, replace=False)\n heights = np.random.uniform(0.5, 1.5, num_peaks)\n for pos, height in zip(positions, heights):\n seq[pos] = height\n return seq\n initial_candidates.append(generate_random_peaks(initial_n))\n \n best_score = -1.0\n best_heights = None\n \n # Try each initial sequence\n for candidate in initial_candidates:\n h = candidate.astype(np.float32).copy()\n h_sum = np.sum(h)\n if h_sum < 0.01:\n h = np.clip(h + (0.01 - h_sum) / initial_n, 0.0, 1000.0)\n \n # Gradient ascent with enhanced exploration\n learning_rate = 0.5 # Increased initial learning rate\n noise_scale_initial = 0.5 # Increased initial noise scale\n max_steps = 300000 # Increased maximum steps\n refine_steps = 500000\n noise_decay = 0.995\n learning_rate_decay = 0.999\n \n start_time = time.time()\n \n # Gradient ascent\n for step in range(max_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 \n noise_scale = noise_scale_initial * (noise_decay ** step)\n learning_rate_current = learning_rate * (learning_rate_decay ** step)\n \n noise = noise_scale * np.random.normal(size=h.shape)\n h = np.clip(h + learning_rate_current * grad_h + noise, 0.0, 1000.0)\n \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 remaining_time = 1000 - (time.time() - start_time)\n if remaining_time < 10:\n break\n \n # Multi-scale refinement\n # First upscale and refine\n h_up = _upsample_1d(h)\n h_up, _ = _single_candidate_finetune(h_up, lr=5e-3, steps=refine_steps)\n h_final = np.clip(h_up, 0.0, 1000.0)\n current_score = evaluate_sequence(h_final.tolist())\n \n # Second upscale and refine\n h_up2 = _upsample_1d(h_final)\n h_up2, _ = _single_candidate_finetune(h_up2, lr=5e-3, steps=refine_steps)\n h_final = np.clip(h_up2, 0.0, 1000.0)\n current_score = evaluate_sequence(h_final.tolist())\n \n if current_score > best_score:\n best_score = current_score\n best_heights = h_final.tolist()\n \n return best_heights\n```",
"env/all/time/policy": 314.8412881754339,
"env/all/time/policy/min": 86.65318298339844,
"env/all/time/policy/max": 430.1409065723419,
"env/all/time/env_step": 2398.673213653732,
"env/all/time/env_step/min": 0.005659341812133789,
"env/all/time/env_step/max": 5008.578496217728,
"env/all/time/reward_compute": 3.0780211091041565e-07,
"env/all/time/reward_compute/min": 2.0116567611694336e-07,
"env/all/time/reward_compute/max": 4.917383193969727e-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": 0.03572501614689827,
"advantage/min": -1.0,
"advantage/max": 15.591583251953125,
"time/assemble_training_data": 10.174549579620361,
"time/kl_vs_base": 92.0185956954956,
"kl_policy_base": 0.0006837469409219921,
"time/train": 595.6562521457672,
"time/save_checkpoint": 13.796337604522705,
"time/total": 6131.06415605545
}
[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