Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac2-step19
ttt-discover-v3-qwen3-8b-ac2-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": 295,
"puct/sampled_size": 8,
"puct/T": 9216,
"puct/scale_last": 0.43846928787632866,
"puct/buffer_value/mean": 0.9233582661376213,
"puct/buffer_value/std": 0.053681100924972434,
"puct/buffer_value/min": 0.5041471954736918,
"puct/buffer_value/max": 0.9426164833500205,
"puct/buffer_timestep/mean": 8.213559322033898,
"puct/buffer_timestep/std": 5.329020567851337,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 17.0,
"puct/buffer_construction_len/mean": 2604.3322033898303,
"puct/buffer_construction_len/std": 2516.0369398840803,
"puct/buffer_construction_len/min": 1024.0,
"puct/buffer_construction_len/max": 32768.0,
"puct/sampled_value/mean": 0.9423427432417322,
"puct/sampled_value/std": 0.00018620251827227357,
"puct/sampled_value/min": 0.942026410997678,
"puct/sampled_value/max": 0.9426164833500205,
"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": 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": 3641.0685908794403,
"env/all/ac_tokens_per_turn": 9358.953125,
"env/all/ob_tokens_per_turn": 3967.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4791784,
"env/all/total_ob_tokens": 2031296,
"env/all/time/sampling_mean": 373.41618882259354,
"env/all/time/sampling_max": 487.5884289741516,
"env/all/time/env_step_mean": 1220.2405654676259,
"env/all/time/env_step_max": 3143.2300930023193,
"env/all/reward/mean": 0.49921342477331804,
"env/all/reward/max": 0.9429107437905139,
"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": 0.49921342477331804,
"env/all/correctness": 0.564453125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.8844196314323143,
"env/all/raw_score/min": 0.19556355492867933,
"env/all/raw_score/max": 0.9429107437905139,
"env/all/initial_raw_score": 0.9423427432417321,
"env/all/initial_raw_score/min": 0.942026410997678,
"env/all/initial_raw_score/max": 0.9426164833500205,
"env/all/msg": "RuntimeError: Program execution failed: TypeError: '>' not supported between instances of 'float' and 'NoneType'",
"env/all/parsed_code": "```python\nimport numpy as np\nimport time\nimport random\nimport math\n\ndef construct_function():\n \"\"\"\n Optimizes the sequence using a combination of global perturbation and local refinement \n with adaptive learning rates and smart initializations. This method improves upon previous \n approaches by introducing smarter initialization strategies, more nuanced learning rates, \n and a mechanism to avoid clamping issues in gradient calculations.\n \"\"\"\n \n def l2_norm_squared(conv, num_points, x_points):\n \"\"\"Calculate the L2 norm squared using Simpson's rule.\"\"\"\n integral = 0.0\n y_points = np.concatenate(([0], conv, [0]))\n x_intervals = np.diff(x_points) # Width of each interval\n for i in range(len(conv) + 1):\n y1 = y_points[i]\n y2 = y_points[i + 1]\n h = x_intervals[i]\n integral += (h / 3.0) * (y1**2 + y1 * y2 + y2**2)\n return integral\n\n def l1_norm(conv, num_points):\n \"\"\"Calculate the L1 norm.\"\"\"\n return (np.sum(np.abs(conv)) / (num_points + 1))\n\n def linf_norm(conv):\n \"\"\"Calculate the L-infinity norm.\"\"\"\n return np.max(np.abs(conv))\n\n def evaluate_sequence(sequence):\n \"\"\"Helper function from the problem.\"\"\"\n # ... [Assuming it's already imported and defined] ...\n\n # --- Initialization Logic ---#\n max_seconds = 1000\n start_time = time.time()\n \n # Use the previous best sequence if available\n if 'height_sequence_1' in globals():\n initial_sequence = np.array(height_sequence_1, dtype=np.float64)\n else:\n # Try multiple initial strategies\n strategies = [\n lambda n: np.random.uniform(0.0, 1000.0, size=n),\n lambda n: np.zeros(n) + 0.1,\n lambda n: np.zeros(n) + 50.0,\n lambda n: np.zeros(n) + 100.0,\n lambda n: np.zeros(n) + 200.0\n ]\n initial_sequence = strategies[0](512)\n initial_sequence = np.clip(initial_sequence, 0.0, 1000.0)\n if np.sum(initial_sequence) < 0.01:\n needed = max(0.01 - np.sum(initial_sequence), 0.01)\n initial_sequence += needed / len(initial_sequence)\n \n best_sequence = initial_sequence.copy()\n best_value = evaluate_sequence(best_sequence.tolist())\n\n # --- Adaptive Optimization Logic ---#\n def update_sequence(seq, lr, grad, clamp_max):\n \"\"\"Apply gradient update with clamping and adjustment.\"\"\"\n seq = seq + lr * grad\n seq = np.clip(seq, 0.0, clamp_max)\n # Check for clamping and adjust learning rate\n clamped = np.any((seq != 0.0) & (seq != 1000.0))\n if not clamped:\n lr *= 0.8 # Reduce learning rate when near optimal point\n return seq, lr\n\n # --- Main Loop ---#\n max_iter = 200000\n learning_rate = 0.5\n lr_decay = 0.9995\n decay_steps = 100\n step_counter = 0\n grad_prev = None\n best_val = best_value\n best_seq = best_sequence.copy()\n\n for step in range(max_iter):\n if time.time() - start_time > max_seconds - 5:\n break\n\n # Random perturbation to escape local minima\n if step % 500 == 0:\n # Try multiple random perturbations\n perturb = np.random.normal(0, 0.05, size=best_seq.shape)\n best_seq = best_seq + perturb\n best_seq = np.clip(best_seq, 0.0, 1000.0)\n \n # Compute the convolution\n conv = np.convolve(best_seq, best_seq, mode='full')\n num_points = len(conv)\n x_points = np.linspace(-0.5, 0.5, num_points + 2)\n\n # Compute the L2 squared norm\n l2_sq = l2_norm_squared(conv, num_points, x_points)\n\n # Compute the L1 norm\n l1 = l1_norm(conv, num_points)\n\n # Compute the L-infinity norm\n linf = linf_norm(conv)\n\n # Check for invalidity (denominator near zero)\n if l1 <= 0.0 or linf <= 0.0:\n # Try restarting with a different initialization\n if 'height_sequence_1' in globals():\n initial_sequence = np.array(height_sequence_1, dtype=np.float64)\n else:\n initial_sequence = np.random.uniform(0.0, 1000.0, size=512)\n initial_sequence = np.clip(initial_sequence, 0.0, 1000.0)\n if np.sum(initial_sequence) < 0.01:\n needed = max(0.01 - np.sum(initial_sequence), 0.01)\n initial_sequence += needed / len(initial_sequence)\n best_sequence = initial_sequence.copy()\n best_value = evaluate_sequence(best_sequence.tolist())\n continue\n\n # Compute the objective function value\n current_value = l2_sq / (l1 * linf)\n if current_value > best_val:\n best_val = current_value\n best_seq = best_sequence.copy()\n print(f\"Found better value: {best_val:.6f} at step {step}\")\n\n # Compute gradient numerically using finite differences to avoid clamping issues\n grad = np.zeros_like(best_seq)\n eps = 1e-5\n for i in range(len(best_seq)):\n # Compute perturbed sequence with one element changed\n pert_seq = best_seq.copy()\n pert_seq[i] += eps\n pert_seq = np.clip(pert_seq, 0.0, 1000.0)\n conv_pert = np.convolve(pert_seq, pert_seq, mode='full')\n l2_sq_pert = l2_norm_squared(conv_pert, len(conv_pert), x_points)\n l1_pert = l1_norm(conv_pert, len(conv_pert))\n linf_pert = linf_norm(conv_pert)\n if l1_pert <= 0.0 or linf_pert <= 0.0:\n grad[i] = 0.0\n else:\n grad[i] = (l2_sq_pert / (l1_pert * linf_pert) - current_value) / eps\n\n # Apply gradient update with adaptive learning rate\n best_seq, learning_rate = update_sequence(best_seq, learning_rate, grad, 1000.0)\n\n # Ensure sum is at least 0.01\n if np.sum(best_seq) < 0.01:\n needed = max(0.01 - np.sum(best_seq), 0.01)\n best_seq += needed / len(best_seq)\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": 373.41618882259354,
"env/all/time/policy/min": 158.06665706634521,
"env/all/time/policy/max": 487.5884289741516,
"env/all/time/env_step": 1220.2405654676259,
"env/all/time/env_step/min": 0.007329463958740234,
"env/all/time/env_step/max": 3143.2300930023193,
"env/all/time/reward_compute": 3.3527612686157227e-07,
"env/all/time/reward_compute/min": 1.8998980522155762e-07,
"env/all/time/reward_compute/max": 6.854534149169922e-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": 0.020607683807611465,
"advantage/min": -1.0,
"advantage/max": 3.706789493560791,
"time/assemble_training_data": 6.608046770095825,
"time/kl_vs_base": 95.66322231292725,
"kl_policy_base": 0.0007291429210454226,
"time/train": 654.0380182266235,
"time/save_checkpoint": 9.169188022613525,
"time/total": 4408.803335189819
}
[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