Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac2-step22
ttt-discover-v3-qwen3-8b-ac2-step22 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 21 (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 21 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 21,
"progress/batch": 21,
"optim/lr": 4e-05,
"progress/done_frac": 0.44,
"puct/buffer_size": 342,
"puct/sampled_size": 8,
"puct/T": 10752,
"puct/scale_last": 0.4389127076413719,
"puct/buffer_value/mean": 0.9253394045539263,
"puct/buffer_value/std": 0.05161736153773509,
"puct/buffer_value/min": 0.5041471954736918,
"puct/buffer_value/max": 0.9430599031150637,
"puct/buffer_timestep/mean": 9.692982456140351,
"puct/buffer_timestep/std": 6.1906275031480575,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 20.0,
"puct/buffer_construction_len/mean": 2698.5438596491226,
"puct/buffer_construction_len/std": 2908.463294156632,
"puct/buffer_construction_len/min": 1024.0,
"puct/buffer_construction_len/max": 32768.0,
"puct/sampled_value/mean": 0.9430254770581256,
"puct/sampled_value/std": 1.8987591472500463e-05,
"puct/sampled_value/min": 0.9430013243579071,
"puct/sampled_value/max": 0.9430599031150637,
"puct/sampled_timestep/mean": 20.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 20.0,
"puct/sampled_timestep/max": 20.0,
"puct/sampled_construction_len/mean": 2048.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 2048.0,
"puct/sampled_construction_len/max": 2048.0,
"time/sampling": 3430.446613550186,
"env/all/ac_tokens_per_turn": 9204.900390625,
"env/all/ob_tokens_per_turn": 3662.875,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4712909,
"env/all/total_ob_tokens": 1875392,
"env/all/time/sampling_mean": 360.82546211453155,
"env/all/time/sampling_max": 467.37548542022705,
"env/all/time/env_step_mean": 1362.1822466757149,
"env/all/time/env_step_max": 2978.1081788539886,
"env/all/reward/mean": 0.5357825400292338,
"env/all/reward/max": 0.9431227831219725,
"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": 0.5357825400292338,
"env/all/correctness": 0.60546875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.8849053564353796,
"env/all/raw_score/min": 0.03328543322318066,
"env/all/raw_score/max": 0.9431227831219725,
"env/all/initial_raw_score": 0.9430254770581254,
"env/all/initial_raw_score/min": 0.9430013243579071,
"env/all/initial_raw_score/max": 0.9430599031150637,
"env/all/msg": "Success; raw_score=0.9308100173737899",
"env/all/parsed_code": "```python\nimport numpy as np\nimport time\nfrom typing import Tuple\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=\"valid\")\n return 2.0 * g_h\n\ndef construct_function():\n \"\"\"\n Optimizes step function sequence using a hybrid multi-scale approach with Adam optimizer and random perturbations.\n Starts with a diverse initialization and scales up iteratively to refine the solution.\n \"\"\"\n target_size = 2048\n max_seconds = 1000\n\n # Initialize sequence\n if 'height_sequence_1' in globals():\n initial_seq = np.array(height_sequence_1, dtype=np.float32)\n if len(initial_seq) < target_size:\n initial_seq = np.pad(initial_seq, (0, target_size - len(initial_seq)), 'constant')\n elif len(initial_seq) > target_size:\n initial_seq = initial_seq[:target_size]\n best_seq = initial_seq.copy()\n else:\n # Generate random initial sequence with sum >= 0.01\n random_seq = np.random.uniform(low=0.0, high=1000.0, size=target_size)\n random_sum = np.sum(random_seq)\n if random_sum < 0.01:\n scaling_factor = 0.01 / random_sum\n random_seq *= scaling_factor\n best_seq = random_seq.copy()\n\n best_value = evaluate_sequence(best_seq.tolist())\n\n start_time = time.time()\n step_counter = 0\n\n max_iter = 100000\n learning_rate = 0.1 # Larger initial learning rate for exploration\n decay_rate = 0.99 # Slightly slower decay\n min_lr = 1e-5\n decay_steps = 50 # More frequent decay for adaptive learning\n\n beta1 = 0.9\n beta2 = 0.999\n epsilon = 1e-8\n\n m = np.zeros_like(best_seq)\n v = np.zeros_like(best_seq)\n\n while time.time() - start_time < max_seconds - 5:\n step_counter += 1\n\n # Compute gradients\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\n \n denom = l1 * linf\n c_value = l2_sq / denom\n\n g_conv = _objective_and_grad_conv(conv)[1]\n grad_h = _grad_h_from_conv_grad(best_seq, g_conv)\n\n # Adam Update\n m = beta1 * m + (1 - beta1) * grad_h\n v = beta2 * v + (1 - beta2) * (grad_h ** 2)\n\n t = step_counter + 1\n m_hat = m / (1 - beta1 ** t)\n v_hat = v / (1 - beta2 ** t)\n\n best_seq = best_seq + learning_rate * m_hat / (np.sqrt(v_hat) + epsilon)\n best_seq = np.clip(best_seq, 0.0, 1000.0)\n\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\n # Decay learning rate\n if step_counter % decay_steps == 0:\n learning_rate *= decay_rate\n learning_rate = max(learning_rate, min_lr)\n\n # Periodically upscale\n if step_counter % 500 == 0:\n # Double the sequence size up to target\n new_size = min(target_size, len(best_seq) * 2)\n if new_size > len(best_seq):\n new_seq = np.repeat(best_seq, 2)[:new_size]\n best_seq = new_seq\n m = np.zeros_like(best_seq)\n v = np.zeros_like(best_seq)\n # Adjust learning rate after upscaling\n learning_rate = 0.05\n # Perturb the sequence with some probability\n if np.random.rand() < 0.05:\n perturb_factor = 0.1\n best_seq += perturb_factor * np.random.randn(*best_seq.shape)\n best_seq = np.clip(best_seq, 0.0, 1000.0)\n\n # Final evaluation\n final_heights = best_seq.tolist()\n final_value = evaluate_sequence(final_heights)\n print(f\"Final C2 lower bound: {final_value:.6f}\")\n return final_heights\n```",
"env/all/time/policy": 360.82546211453155,
"env/all/time/policy/min": 85.67169904708862,
"env/all/time/policy/max": 467.37548542022705,
"env/all/time/env_step": 1362.1822466757149,
"env/all/time/env_step/min": 0.00643610954284668,
"env/all/time/env_step/max": 2978.1081788539886,
"env/all/time/reward_compute": 2.4260953068733215e-07,
"env/all/time/reward_compute/min": 1.6391277313232422e-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": 0.0197131410241127,
"advantage/min": -1.0,
"advantage/max": 3.0460753440856934,
"time/assemble_training_data": 6.353882074356079,
"time/kl_vs_base": 96.485360622406,
"kl_policy_base": 0.0007516420446336269,
"time/train": 626.0405163764954,
"time/save_checkpoint": 11.622063159942627,
"time/total": 4173.631470680237
}
[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