Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step36
ttt-discover-v3-qwen3-8b-ac1-step36 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 ac1. Checkpoint saved after training step 35 (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 ac1. Checkpoint saved
after training step 35 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 35,
"progress/batch": 35,
"optim/lr": 4e-05,
"progress/done_frac": 0.72,
"puct/buffer_size": 568,
"puct/sampled_size": 8,
"puct/T": 17920,
"puct/scale_last": 0.4936642696442801,
"puct/buffer_value/mean": -1.5194106473872442,
"puct/buffer_value/std": 0.07001531636796689,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5063357303557203,
"puct/buffer_timestep/mean": 16.746478873239436,
"puct/buffer_timestep/std": 10.249999395203218,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 34.0,
"puct/buffer_construction_len/mean": 1076.6390845070423,
"puct/buffer_construction_len/std": 563.0577594835675,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.506338234710569,
"puct/sampled_value/std": 2.504354883477311e-06,
"puct/sampled_value/min": -1.5063407397668263,
"puct/sampled_value/max": -1.5063357303557203,
"puct/sampled_timestep/mean": 34.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 34.0,
"puct/sampled_timestep/max": 34.0,
"puct/sampled_construction_len/mean": 1000.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 1000.0,
"puct/sampled_construction_len/max": 1000.0,
"time/sampling": 5306.268406152725,
"env/all/ac_tokens_per_turn": 9063.556640625,
"env/all/ob_tokens_per_turn": 3061.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4640541,
"env/all/total_ob_tokens": 1567552,
"env/all/time/sampling_mean": 551.0341132110916,
"env/all/time/sampling_max": 827.357837677002,
"env/all/time/env_step_mean": 2298.1673776060343,
"env/all/time/env_step_max": 4485.868552684784,
"env/all/reward/mean": 0.6175360384618072,
"env/all/reward/max": 0.6638626258471795,
"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": 0.6175360384618072,
"env/all/correctness": 0.94921875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.5559946856089772,
"env/all/raw_score/min": 1.5063357303557148,
"env/all/raw_score/max": 4.449189930142729,
"env/all/initial_raw_score": -1.5063382347105692,
"env/all/initial_raw_score/min": -1.5063407397668263,
"env/all/initial_raw_score/max": -1.5063357303557203,
"env/all/msg": "Success; raw_score=1.5063357303557203",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy.optimize import linprog\nimport random\nimport copy\n\ndef propose_candidate(seed=42, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score by \n combining adaptive constraint selection, dynamic perturbation, and efficient optimization.\n The method avoids local minima by exploring diverse solutions and refines the search through careful \n constraint sampling and gradient-aware updates.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + 1000 - 10 # Run for up to 1000 seconds\n\n # Use the existing best sequence if available\n if 'height_sequence_1' in globals() and isinstance(height_sequence_1, (list, np.ndarray)):\n best_sequence = np.array(list(height_sequence_1))\n else:\n # Initialize with a uniform distribution for broader search\n n = 1000\n best_sequence = np.ones(n) * np.sqrt(2 * n) / n\n\n curr_sequence = best_sequence.copy()\n best_score = evaluate_sequence(curr_sequence.tolist())\n no_improvement_counter = 0\n iteration_count = 0\n\n # Parameters for line search\n k_values = [1500, 1200, 900, 600, 300] # More aggressive constraint selection\n\n while time.time() < deadline:\n current_time = time.time()\n try:\n # Compute current convolution and max_b\n conv = np.convolve(curr_sequence, curr_sequence)\n max_b = np.max(conv)\n sum_a = np.sum(curr_sequence)\n if sum_a < 0.001:\n raise ValueError(\"Sum too small\")\n\n # Select constraints dynamically based on iteration\n k = max(200, 1500 - iteration_count * 1)\n sorted_conv = sorted(enumerate(conv), key=lambda x: -x[1])\n top_indices = [i for i, _ in sorted_conv[:k]]\n # Add edge and spaced-out indices for coverage\n edge_indices = list(range(0, len(conv), 500))\n spaced_indices = list(range(0, len(conv), 100))\n top_indices = np.unique(np.concatenate([top_indices, edge_indices, spaced_indices])).tolist()\n A = []\n for j in top_indices:\n if j < 0 or j >= len(conv): # Ensure valid indices\n continue\n row = np.zeros(n)\n for m in range(n):\n i = j - m\n if 0 <= i < n:\n row[m] = curr_sequence[i]\n A.append(row)\n A = np.array(A)\n b = np.full(len(A), max_b - max_b * 0.00001) # Slight relax for feasibility\n\n # Define LP problem: maximize sum of g_0 subject to constraints\n c = [-1.0] * n\n bounds = [(0.0, 1000.0) for _ in range(n)]\n\n # Solve the LP using 'highs' for performance\n res = linprog(c=c, A_ub=A, b_ub=b, bounds=bounds, method='highs', options={\"feastol\": 1e-8})\n if res.success:\n g_0 = res.x\n sum_g = np.sum(g_0)\n if sum_g == 0:\n continue # Avoid division by zero\n # Normalize g_0 to have the same sum as curr_sequence\n g_0_normalized = g_0 / sum_g * sum_a\n\n # Perform adaptive line search with increased iterations\n def find_optimal_alpha(curr_seq, g0_norm):\n low = 0.0\n high = 1.0\n for _ in range(150): # More iterations for convergence\n m1 = low + (high - low)/3\n m2 = high - (high - low)/3\n new_seq1 = (1 - m1) * curr_seq + m1 * g0_norm\n new_seq2 = (1 - m2) * curr_seq + m2 * g0_norm\n # Calculate new max_b\n conv_new1 = np.convolve(new_seq1, new_seq1)\n conv_new2 = np.convolve(new_seq2, new_seq2)\n new_max_b1 = np.max(conv_new1)\n new_max_b2 = np.max(conv_new2)\n \n if new_max_b1 < new_max_b2:\n high = m2\n else:\n low = m1\n alpha_opt = (low + high) / 2\n new_seq = (1 - alpha_opt) * curr_seq + alpha_opt * g0_norm\n return new_seq, np.max(np.convolve(new_seq, new_seq))\n \n new_sequence, new_score = find_optimal_alpha(curr_sequence, g_0_normalized)\n \n # Relax the max_b check to allow more updates\n conv_new = np.convolve(new_sequence, new_sequence)\n new_max_b = np.max(conv_new)\n\n curr_sequence = new_sequence\n\n # Evaluate and update best sequence\n curr_score = evaluate_sequence(curr_sequence.tolist())\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = curr_sequence.copy()\n no_improvement_counter = 0\n print(f\"New best score: {best_score}\")\n else:\n no_improvement_counter += 1\n else:\n print(\"LP solution not found, breaking\")\n break\n\n # Intelligent perturbation to avoid local minima\n if iteration_count % 20 == 0: # Reduced frequency\n # Targeted perturbation on peak regions\n peaks = np.argsort(conv)[-50:]\n perturbation_indices = np.random.choice(peaks, size=50, replace=False)\n perturbation = np.zeros(n)\n perturbation[perturbation_indices] = np.random.normal(0, 0.01, 50)\n perturbation = np.clip(perturbation, -0.05, 0.05)\n curr_sequence += perturbation\n curr_sequence = np.maximum(0.0, curr_sequence)\n # Allow sum to vary, only enforce minimum threshold\n curr_sequence = np.clip(curr_sequence, 0.0, 1000.0)\n curr_sequence = curr_sequence / np.sum(curr_sequence) * np.sqrt(2 * n)\n\n # Early stopping if no improvement for many iterations\n iteration_count += 1\n if iteration_count % 100 == 0:\n print(f\"Iteration {iteration_count}, current best score: {best_score}\")\n if no_improvement_counter > 200: # Increased threshold\n print(\"No improvement for 200 iterations, stopping early\")\n break\n\n except Exception as e:\n print(f\"Error during optimization: {e}. Skipping update.\")\n continue\n\n # Final safety check to ensure non-negative and valid values\n final_sequence = np.maximum(0.0, best_sequence).tolist()\n return [float(x) for x in final_sequence]\n```",
"env/all/time/policy": 551.0341132110916,
"env/all/time/policy/min": 250.6163845062256,
"env/all/time/policy/max": 827.357837677002,
"env/all/time/env_step": 2298.1673776060343,
"env/all/time/env_step/min": 2.2614247798919678,
"env/all/time/env_step/max": 4485.868552684784,
"env/all/time/reward_compute": 3.390014171600342e-07,
"env/all/time/reward_compute/min": 2.421438694000244e-07,
"env/all/time/reward_compute/max": 4.954636096954346e-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": 0.003166562644764781,
"advantage/min": -1.0,
"advantage/max": 0.36956512928009033,
"time/assemble_training_data": 10.500015497207642,
"time/kl_vs_base": 150.55863976478577,
"kl_policy_base": 0.0006678792415186763,
"time/train": 1176.003023147583,
"time/save_checkpoint": 21.531810760498047,
"time/total": 6667.71035194397
}
[2026-07-09T06:24:18+00:00] job=1812628 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:30:32+00:00] job=1813127 node=node-22 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:07:08+00:00] job=1813128 node=node-6 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-11T13:13:37+00:00] job=1824165 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=yes