Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac2-step26
ttt-discover-v3-qwen3-8b-ac2-step26 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 25 (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 25 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 25,
"progress/batch": 25,
"optim/lr": 4e-05,
"progress/done_frac": 0.52,
"puct/buffer_size": 406,
"puct/sampled_size": 8,
"puct/T": 12800,
"puct/scale_last": 0.4393783031691666,
"puct/buffer_value/mean": 0.9281459018855367,
"puct/buffer_value/std": 0.047816801184062174,
"puct/buffer_value/min": 0.5041471954736918,
"puct/buffer_value/max": 0.9435254986428584,
"puct/buffer_timestep/mean": 11.711822660098521,
"puct/buffer_timestep/std": 7.366092475172098,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 24.0,
"puct/buffer_construction_len/mean": 2616.1724137931033,
"puct/buffer_construction_len/std": 2683.410780806649,
"puct/buffer_construction_len/min": 1024.0,
"puct/buffer_construction_len/max": 32768.0,
"puct/sampled_value/mean": 0.9432782933979078,
"puct/sampled_value/std": 0.0001274471835263143,
"puct/sampled_value/min": 0.9431995196146987,
"puct/sampled_value/max": 0.9435254986428584,
"puct/sampled_timestep/mean": 24.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 24.0,
"puct/sampled_timestep/max": 24.0,
"puct/sampled_construction_len/mean": 2560.0,
"puct/sampled_construction_len/std": 886.8100134752651,
"puct/sampled_construction_len/min": 2048.0,
"puct/sampled_construction_len/max": 4096.0,
"time/sampling": 4011.8252437114716,
"env/all/ac_tokens_per_turn": 9814.806640625,
"env/all/ob_tokens_per_turn": 4182.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 5025181,
"env/all/total_ob_tokens": 2141248,
"env/all/time/sampling_mean": 409.25440851645544,
"env/all/time/sampling_max": 516.8386814594269,
"env/all/time/env_step_mean": 1609.307591218967,
"env/all/time/env_step_max": 3494.348457813263,
"env/all/reward/mean": 0.6590073566919118,
"env/all/reward/max": 0.9436554622718666,
"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": 0.6590073566919118,
"env/all/correctness": 0.71875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.9168798006148337,
"env/all/raw_score/min": 0.42857604837892244,
"env/all/raw_score/max": 0.9436554622718666,
"env/all/initial_raw_score": 0.9432782933979078,
"env/all/initial_raw_score/min": 0.9431995196146987,
"env/all/initial_raw_score/max": 0.9435254986428584,
"env/all/msg": "Success; raw_score=0.9431401726038015",
"env/all/parsed_code": "```python\nimport numpy as np\nimport time\nfrom typing import Tuple\nimport random\n\ndef _simpson_l2sq(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n \"\"\"Compute ||f*f||_2^2 via Simpson's rule with endpoint zeros and gradient.\"\"\"\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 \"\"\"Compute ||f*f||_1 and its gradient.\"\"\"\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 \"\"\"Compute ||f*f||_inf and its subgradient.\"\"\"\n if conv.size == 0:\n return 0.0, np.zeros_like(conv)\n m = float(np.max(conv))\n mask = conv == m\n count = int(mask.sum())\n if count == 0 or m <= 0.0:\n return m, np.zeros_like(conv)\n grad = mask.astype(conv.dtype)\n return m, grad\n\ndef _objective_and_grad_conv(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n \"\"\"Compute C = l2_sq / (l1 * linf) and its gradient.\"\"\"\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 \"\"\"Compute gradient of C w.r.t h from gradient of C w.r.t conv.\"\"\"\n h_rev = h[::-1]\n g_h = np.convolve(g_conv, h_rev, mode=\"full\")\n N = len(h)\n # Extract the central N elements for the gradient\n gradient = g_h[N-1 : N-1 + N]\n return 2.0 * gradient\n\ndef construct_function():\n \"\"\"\n Optimizes step function sequence using a fine-tuned gradient ascent strategy with better constraints handling.\n \n This function employs a combination of multi-scale initialization and adaptive gradient ascent, \n with improved learning rate control and sum normalization. The optimization is performed with\n the aim to push the lower bound of the score closer to the ideal value.\n \"\"\"\n # Parameters for the algorithm\n n_start = 64 # Initial scale for the function\n max_seconds = 1000 # Time limit in seconds\n num_initializations = 8 # Diverse initializations to prevent local minima\n target_length = 4096 # Target resolution for the final sequence\n learning_rate = 0.01 # Tuned to prevent overshooting in gradient steps\n decay_rate = 0.999 # Slow learning rate decay for better convergence\n min_lr = 1e-5 # Minimum learning rate\n decay_steps = 100 # Steps after which to decay the learning rate\n beta1 = 0.9\n beta2 = 0.999\n epsilon = 1e-8 # Small epsilon for numerical stability\n\n # Initialize with previous best if possible\n initial_sequences = []\n if 'height_sequence_1' in globals():\n initial_sequences.append(np.array(height_sequence_1, dtype=np.float32))\n \n # Generate various structured initial sequences\n for _ in range(num_initializations):\n # Flat uniform distribution\n initial_sequences.append(np.full(n_start, (0.01 + 1e-6) / n_start, dtype=np.float32))\n\n # Random peak initialization\n seq = np.zeros(n_start, dtype=np.float32)\n num_random_peaks = 8\n peak_height = (0.01 + 1e-6) / num_random_peaks\n indices = np.random.choice(n_start, num_random_peaks, replace=False)\n seq[indices] = peak_height\n initial_sequences.append(seq)\n\n # Single peak\n seq = np.zeros(n_start, dtype=np.float32)\n seq[n_start // 2] = 0.01 + 1e-6\n initial_sequences.append(seq)\n\n # Sinusoidal pattern\n seq = np.sin(np.linspace(0, 2 * np.pi, n_start))\n seq = np.clip(seq, 0.0, 1.0)\n sum_val = np.sum(seq)\n if sum_val > 0:\n seq = seq * (0.01 + 1e-6) / sum_val\n else:\n seq = np.full(n_start, (0.01 + 1e-6) / n_start)\n initial_sequences.append(seq)\n\n # Gaussian-like pattern\n seq = np.zeros(n_start, dtype=np.float32)\n center = n_start // 2\n std = n_start // 8\n for i in range(n_start):\n seq[i] = np.exp(-((i - center) ** 2) / (2 * std ** 2))\n sum_val = np.sum(seq)\n if sum_val > 0:\n seq = seq * (0.01 + 1e-6) / sum_val\n else:\n seq = np.full(n_start, (0.01 + 1e-6) / n_start)\n initial_sequences.append(seq)\n\n # Alternating pattern\n seq = np.zeros(n_start, dtype=np.float32)\n total_sum = 0.01 + 1e-6\n m = n_start // 2\n a = total_sum * 0.6 / m\n b = total_sum * 0.4 / m\n for i in range(n_start):\n if i % 2 == 0:\n seq[i] = a\n else:\n seq[i] = b\n initial_sequences.append(seq)\n\n # Evaluate initial sequences\n best_seq = None\n best_value = -float('inf')\n for seq in initial_sequences:\n try:\n current_value = evaluate_sequence(seq.tolist())\n except Exception as e:\n print(f\"Error evaluating initial sequence: {e}\")\n continue\n if current_value > best_value:\n best_value = current_value\n best_seq = seq.copy()\n\n # Fallback to random initialization if needed\n if best_seq is None:\n best_seq = np.random.rand(n_start) * 1000.0\n best_seq = np.clip(best_seq, 0.0, 1000.0)\n best_seq = best_seq * ((0.01 + 1e-6) / np.sum(best_seq))\n best_value = evaluate_sequence(best_seq.tolist())\n\n # Upsample to target length\n upsample_factor = (target_length + len(best_seq) - 1) // len(best_seq)\n upsampled_seq = np.repeat(best_seq, upsample_factor, axis=0)\n upsampled_seq = upsampled_seq[:target_length]\n best_seq = upsampled_seq\n best_value = evaluate_sequence(best_seq.tolist())\n\n start_time = time.time()\n best_seq = best_seq.copy()\n best_value = evaluate_sequence(best_seq.tolist())\n\n # Adam optimizer parameters\n m = np.zeros_like(best_seq)\n v = np.zeros_like(best_seq)\n t = 0\n\n # Optimization loop\n for _ in range(100000):\n step_counter = _ + 1\n\n # Compute gradient\n conv = np.convolve(best_seq, best_seq, mode='full')\n l2_sq, g_l2 = _simpson_l2sq(conv)\n l1, g_l1 = _l1(conv)\n linf, g_linf = _linf(conv)\n\n if l1 <= 0.0 or linf <= 0.0:\n break # Avoid division by zero\n\n denom = l1 * linf\n c_value = l2_sq / denom\n\n # Compute gradient with respect to convolution\n g_conv = _objective_and_grad_conv(conv)[1]\n\n # Compute gradient with respect to h\n grad_h = _grad_h_from_conv_grad(best_seq, g_conv)\n\n # Adam update\n t += 1\n m = beta1 * m + (1 - beta1) * grad_h\n v = beta2 * v + (1 - beta2) * (grad_h ** 2)\n m_hat = m / (1 - beta1 ** t)\n v_hat = v / (1 - beta2 ** t)\n step = learning_rate * m_hat / (np.sqrt(v_hat) + epsilon)\n best_seq = best_seq + step\n\n # Apply clipping and sum normalization\n best_seq = np.clip(best_seq, 0.0, 1000.0)\n sum_seq = np.sum(best_seq)\n if sum_seq < 0.01 + 1e-6:\n scale_factor = (0.01 + 1e-6) / sum_seq\n best_seq *= scale_factor\n best_seq = np.clip(best_seq, 0.0, 1000.0)\n\n # Track best value and check for stopping condition\n current_value = evaluate_sequence(best_seq.tolist())\n if current_value > best_value:\n best_value = current_value\n best_seq = best_seq.copy()\n if time.time() - start_time > max_seconds - 5:\n break\n if step_counter % decay_steps == 0:\n learning_rate *= decay_rate\n\n # Final evaluation\n final_value = evaluate_sequence(best_seq.tolist())\n print(f\"Final C2 Lower Bound: {final_value:.6f}\")\n return best_seq.tolist()\n```",
"env/all/time/policy": 409.25440851645544,
"env/all/time/policy/min": 73.78286600112915,
"env/all/time/policy/max": 516.8386814594269,
"env/all/time/env_step": 1609.307591218967,
"env/all/time/env_step/min": 0.006414890289306641,
"env/all/time/env_step/max": 3494.348457813263,
"env/all/time/reward_compute": 2.7241185307502747e-07,
"env/all/time/reward_compute/min": 1.7881393432617188e-07,
"env/all/time/reward_compute/max": 4.880130290985107e-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": 0.01764158345758915,
"advantage/min": -1.0,
"advantage/max": 2.549913167953491,
"time/assemble_training_data": 10.996619701385498,
"time/kl_vs_base": 107.27001857757568,
"kl_policy_base": 0.0006685287808068097,
"time/train": 692.8661391735077,
"time/save_checkpoint": 22.21803617477417,
"time/total": 4848.634025335312
}
[2026-07-09T06:38:34+00:00] job=1812632 node=node-30 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T06:45:52+00:00] job=1812704 node=node-30 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:00:42+00:00] job=1812735 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:26:33+00:00] job=1812827 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:21:09+00:00] job=1813131 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T14:53:48+00:00] job=1813132 node=node-2 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-10T03:31:51+00:00] job=1816627 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-10T03:54:03+00:00] job=1816628 node=node-29 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-10T07:56:44+00:00] job=1817463 node=node-27 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-10T09:10:26+00:00] job=1817464 node=node-7 ngpu=6 ntrain=2 replicas=4 flash_attn=yes