Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step5
ttt-discover-v3-qwen3-8b-ac1-s2-step5 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 4 (0-indexed). Strict upstream eval parity: 1100s hard k…
Downloads · 30 days
0
Access
Public
Updated Jul 9, 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 4 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 4,
"progress/batch": 4,
"optim/lr": 4e-05,
"progress/done_frac": 0.1,
"puct/buffer_size": 72,
"puct/sampled_size": 8,
"puct/T": 2048,
"puct/scale_last": 0.49062007027572463,
"puct/buffer_value/mean": -1.622073227139096,
"puct/buffer_value/std": 0.18251527045070065,
"puct/buffer_value/min": -2.000000000000007,
"puct/buffer_value/max": -1.5076404096987763,
"puct/buffer_timestep/mean": 1.2222222222222223,
"puct/buffer_timestep/std": 1.314684396244359,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 3.0,
"puct/buffer_construction_len/mean": 1829.3055555555557,
"puct/buffer_construction_len/std": 1782.7264983016642,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5112542465286445,
"puct/sampled_value/std": 0.0024279869954351462,
"puct/sampled_value/min": -1.5136562540637661,
"puct/sampled_value/max": -1.5076404096987763,
"puct/sampled_timestep/mean": 3.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 3.0,
"puct/sampled_timestep/max": 3.0,
"puct/sampled_construction_len/mean": 1017.5,
"puct/sampled_construction_len/std": 30.31088913245535,
"puct/sampled_construction_len/min": 1000.0,
"puct/sampled_construction_len/max": 1070.0,
"time/sampling": 5322.525642633438,
"env/all/ac_tokens_per_turn": 7726.9296875,
"env/all/ob_tokens_per_turn": 3005.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 3956188,
"env/all/total_ob_tokens": 1538752,
"env/all/time/sampling_mean": 262.73317474033684,
"env/all/time/sampling_max": 338.64862084388733,
"env/all/time/env_step_mean": 2688.9616164914332,
"env/all/time/env_step_max": 4977.846259593964,
"env/all/reward/mean": 0.5362513075368255,
"env/all/reward/max": 0.6633035181856087,
"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": 0.5362513075368255,
"env/all/correctness": 0.86328125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 2.0619498117057997,
"env/all/raw_score/min": 1.5076054414762505,
"env/all/raw_score/max": 145.0075284611655,
"env/all/initial_raw_score": -1.5112542465286443,
"env/all/initial_raw_score/min": -1.5136562540637661,
"env/all/initial_raw_score/max": -1.5076404096987763,
"env/all/msg": "Success; raw_score=1.5076385989635084",
"env/all/parsed_code": "```python\n\"\"\"Enhanced Adaptive Convolution Minimization with Optimized LP and Targeted Perturbation\"\"\"\n\nimport time\nimport numpy as np\nfrom scipy import optimize\nimport copy\nimport random\n\nlinprog = optimize.linprog\n\ndef get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using LP with optimized constraint selection and enhanced line search.\"\"\"\n n = len(sequence)\n if n == 0:\n return None\n\n sum_seq = np.sum(sequence)\n if sum_seq <= 0.0:\n return None\n\n # Normalize sequence to a fixed sum for LP optimization\n normalized_seq = [x * np.sqrt(2 * n) / sum_seq for x in sequence]\n conv = np.convolve(normalized_seq, normalized_seq)\n max_b = np.max(conv)\n\n # Select top K positions with highest convolution values\n top_k = min(200, 2 * n - 1) # Increased to 200 for tighter constraints\n relevant_positions = np.argsort(conv)[::-1][:top_k]\n\n # Solve LP with extended time limit and more constraints\n g_fun = solve_convolution_lp(normalized_seq, max_b, relevant_positions)\n if g_fun is None:\n return None\n\n sum_g = np.sum(g_fun)\n if sum_g <= 0.0:\n return None\n\n # Normalize g_fun to align with the initial normalization\n normalized_g = [x * np.sqrt(2 * n) / sum_g for x in g_fun]\n\n # Enhanced line search with larger range and more iterations\n best_t = 0.0\n best_score = float('inf')\n t_low, t_high = 0.0, 1.0 # Expanded range to allow more drastic improvements\n for _ in range(100): # Increased number of iterations for better convergence\n t = (t_low + t_high) / 2\n new_sequence = [(1 - t) * x + t * y for x, y in zip(sequence, normalized_g)]\n try:\n curr_score = evaluate_sequence(new_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_t = t\n if curr_score < best_score * 0.9: # Aggressive refinement\n t_low = best_t\n else:\n t_high = best_t\n else:\n t_high = best_t\n except Exception:\n break\n return [(1 - best_t) * x + best_t * y for x, y in zip(sequence, normalized_g)]\n\ndef solve_convolution_lp(f_sequence, rhs, relevant_positions):\n \"\"\"Solves the LP with optimized constraint selection and efficient solver.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n # Compute convolution\n conv = np.convolve(f_sequence, f_sequence)\n max_conv = np.max(conv)\n \n # Build LP constraints\n a_ub = []\n b_ub = []\n for k in relevant_positions:\n row = np.zeros(n)\n for i in range(n):\n j = k - i\n if 0 <= j < n:\n row[j] = f_sequence[i]\n a_ub.append(row)\n b_ub.append(rhs)\n\n # Non-negativity constraints\n a_ub_nonneg = -np.eye(n)\n b_ub_nonneg = np.zeros(n)\n a_ub = np.vstack([a_ub, a_ub_nonneg])\n b_ub = np.hstack([b_ub, b_ub_nonneg])\n\n # Solve with optimized solver with extended time limit\n result = linprog(\n c=-np.ones(n),\n A_ub=a_ub,\n b_ub=b_ub,\n bounds=[(0, 1000.0) for _ in range(n)],\n method='highs', # Use high-performance solver\n options={\n \"time_limit\": 60.0, # Increased time limit for LP\n \"disp\": False,\n },\n )\n if result.success:\n return result.x\n return None\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n \n # Initialize with height_sequence_1 if available\n best_sequence = []\n if 'height_sequence_1' in globals():\n best_sequence = list(np.asarray(height_sequence_1, dtype=float))\n else:\n # Fallback to structured initialization using exponential decay\n n = 1000\n base = np.exp(-0.1 * np.arange(n)) * 0.5\n best_sequence = [max(0.0, base[i]) for i in range(n)]\n \n current_sequence = best_sequence.copy()\n best_score = evaluate_sequence(best_sequence)\n \n # Adaptive Perturbation Strategy\n def targeted_perturb(seq):\n conv = np.convolve(seq, seq)\n max_positions = np.argsort(conv)[::-1][:50] # Top 50 positions\n involved_elements = set()\n for k in max_positions:\n for i in range(len(seq)):\n j = k - i\n if 0 <= j < len(seq):\n involved_elements.add(i)\n involved_elements.add(j)\n if not involved_elements:\n return perturb_sequence(seq)\n idx = np.random.choice(list(involved_elements))\n perturbation = np.random.normal(0, 0.1) * 0.5\n new_val = max(0.0, seq[idx] + perturbation)\n return [new_val if i == idx else seq[i] for i in range(len(seq))]\n \n def perturb_sequence(seq):\n idx = np.random.choice(len(seq))\n perturbation = np.random.normal(0, 0.02) * 0.5\n new_val = max(0.0, seq[idx] + perturbation)\n return [new_val if i == idx else seq[i] for i in range(len(seq))]\n \n while time.time() < deadline:\n h_function = get_good_direction_to_move_into(current_sequence)\n if h_function is None:\n # Use targeted perturbation to avoid getting stuck\n current_sequence = targeted_perturb(current_sequence)\n else:\n current_sequence = h_function\n\n try:\n curr_score = evaluate_sequence(current_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = current_sequence.copy()\n print(f\"New best: {best_score}\")\n except Exception:\n pass\n\n return [float(max(0.0, x)) for x in best_sequence]\n```",
"env/all/time/policy": 262.73317474033684,
"env/all/time/policy/min": 96.99990105628967,
"env/all/time/policy/max": 338.64862084388733,
"env/all/time/env_step": 2688.9616164914332,
"env/all/time/env_step/min": 0.007209062576293945,
"env/all/time/env_step/max": 4977.846259593964,
"env/all/time/reward_compute": 3.166496753692627e-07,
"env/all/time/reward_compute/min": 2.1979212760925293e-07,
"env/all/time/reward_compute/max": 6.891787052154541e-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": 0.035037923604249954,
"advantage/min": -1.0,
"advantage/max": 23.644329071044922,
"time/assemble_training_data": 8.928799629211426,
"time/kl_vs_base": 89.52173709869385,
"kl_policy_base": 0.000641676306258887,
"time/train": 507.30142736434937,
"time/save_checkpoint": 15.307801008224487,
"time/total": 5945.353511333466
}
[2026-07-09T06:24:18+00:00] job=1812630 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T09:14:17+00:00] job=1813129 node=node-28 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:46:36+00:00] job=1813130 node=node-22 ngpu=6 ntrain=2 replicas=4 flash_attn=yes