Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step39
ttt-discover-v3-qwen3-8b-ac1-step39 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 38 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 12, 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 38 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 38,
"progress/batch": 38,
"optim/lr": 4e-05,
"progress/done_frac": 0.78,
"puct/buffer_size": 616,
"puct/sampled_size": 8,
"puct/T": 19456,
"puct/scale_last": 0.5401118663276994,
"puct/buffer_value/mean": -1.5209395678215962,
"puct/buffer_value/std": 0.07422473900988209,
"puct/buffer_value/min": -2.0464450913574384,
"puct/buffer_value/max": -1.506333225029739,
"puct/buffer_timestep/mean": 18.246753246753247,
"puct/buffer_timestep/std": 11.115866563547502,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 37.0,
"puct/buffer_construction_len/mean": 1070.6672077922078,
"puct/buffer_construction_len/std": 541.065712421733,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.506335104153439,
"puct/sampled_value/std": 1.0846141902475445e-06,
"puct/sampled_value/min": -1.5063357303557168,
"puct/sampled_value/max": -1.506333225029739,
"puct/sampled_timestep/mean": 37.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 37.0,
"puct/sampled_timestep/max": 37.0,
"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": 3398.044417858124,
"env/all/ac_tokens_per_turn": 9373.900390625,
"env/all/ob_tokens_per_turn": 2772.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4799437,
"env/all/total_ob_tokens": 1419584,
"env/all/time/sampling_mean": 573.7303740074858,
"env/all/time/sampling_max": 829.7616763114929,
"env/all/time/env_step_mean": 1146.3991196984425,
"env/all/time/env_step_max": 2579.623235464096,
"env/all/reward/mean": 0.571007731097346,
"env/all/reward/max": 0.6638637299802559,
"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": 0.571007731097346,
"env/all/correctness": 0.888671875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 5.961073180583858,
"env/all/raw_score/min": 1.5063332250296366,
"env/all/raw_score/max": 2000.0,
"env/all/initial_raw_score": -1.5063351041534387,
"env/all/initial_raw_score/min": -1.5063357303557168,
"env/all/initial_raw_score/max": -1.506333225029739,
"env/all/msg": "Success; raw_score=1.506333225029739",
"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 \n by dynamically adjusting constraints, using a more efficient line search, and \n allowing flexible normalization. This method avoids local minima by balancing\n exploration and exploitation, and leverages golden section search for line search.\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 and flexible normalization\n n = 500 # Testing lower length for faster convergence\n initial_sequence = np.ones(n)\n initial_sequence = np.clip(initial_sequence, 0.0, 1000.0)\n best_sequence = initial_sequence.copy()\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 = len(curr_sequence)\n\n # Parameters for line search\n k_values = [200, 300, 500] # Focus on critical convolution points\n perturbation_steps = 200 # Perturbation frequency\n perturbation_scale = 0.01 # Smaller perturbations for steady refinement\n alpha_search_iterations = 50 # Reduce iterative overhead\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 based on convolution indices\n k = k_values[iteration_count % len(k_values)]\n sorted_conv = sorted(enumerate(conv), key=lambda x: -x[1])\n top_indices = [i for i, _ in sorted_conv[:k]]\n spaced_indices = list(range(2, len(conv), 150))\n all_indices = np.unique(np.concatenate([top_indices, spaced_indices])).tolist()\n A = []\n for j in all_indices:\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(all_indices), max_b) # Dynamic epsilon adjustment\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\n # No normalization, keep sum_a flexible\n g_0_normalized = g_0.copy()\n\n # Perform golden section search to minimize the evaluation score\n # Define function to minimize: evaluate_sequence after blending with g_0\n def objective(alpha):\n new_sequence = (1 - alpha) * curr_sequence + alpha * g_0_normalized\n new_sequence = np.clip(new_sequence, 0.0, 1000.0)\n return evaluate_sequence(new_sequence.tolist())\n\n # Golden section search between 0 and 1\n phi = (5 ** 0.5 - 1) / 2 # ~0.618\n low, high = 0.0, 1.0\n for _ in range(50):\n m1 = high - phi * (high - low)\n m2 = low + phi * (high - low)\n f1 = objective(m1)\n f2 = objective(m2)\n if f1 < f2:\n high = m2\n else:\n low = m1\n alpha_opt = (low + high) / 2\n new_sequence = (1 - alpha_opt) * curr_sequence + alpha_opt * g_0_normalized\n\n # Ensure non-negative and bounded values\n new_sequence = np.clip(new_sequence, 0.0, 1000.0)\n\n # Evaluate and update best sequence\n curr_score = evaluate_sequence(new_sequence.tolist())\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = new_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 # Targeted perturbation: focus on high-impact areas\n if iteration_count % perturbation_steps == 0:\n # Identify indices with high contribution to max_b\n conv = np.convolve(curr_sequence, curr_sequence)\n top_k = int(n / 10)\n top_indices = np.argsort(conv)[-top_k:]\n bottom_k = int(n / 10)\n bottom_indices = np.argsort(conv)[:bottom_k]\n \n # Collect original indices contributing to these top_k and bottom_k convolution indices\n peak_indices = set()\n valley_indices = set()\n for k in top_indices:\n start = max(0, k - (n - 1))\n end = min(n - 1, k)\n for i in range(start, end + 1):\n peak_indices.add(i)\n for k in bottom_indices:\n start = max(0, k - (n - 1))\n end = min(n - 1, k)\n for i in range(start, end + 1):\n valley_indices.add(i)\n peak_indices = list(peak_indices)\n valley_indices = list(valley_indices)\n if not peak_indices:\n continue\n if not valley_indices:\n continue\n \n # Perturb peaks by reducing and valleys by increasing\n if len(peak_indices) > 0:\n curr_sequence[peak_indices] -= np.random.uniform(0.0001, 0.001, len(peak_indices))\n curr_sequence[peak_indices] = np.maximum(curr_sequence[peak_indices], 0.0)\n if len(valley_indices) > 0:\n curr_sequence[valley_indices] += np.random.uniform(0.0001, 0.001, len(valley_indices))\n curr_sequence[valley_indices] = np.minimum(curr_sequence[valley_indices], 1000.0)\n\n # Ensure non-negativity and bounded values\n curr_sequence = np.clip(curr_sequence, 0.0, 1000.0)\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": 573.7303740074858,
"env/all/time/policy/min": 182.787428855896,
"env/all/time/policy/max": 829.7616763114929,
"env/all/time/env_step": 1146.3991196984425,
"env/all/time/env_step/min": 0.006239175796508789,
"env/all/time/env_step/max": 2579.623235464096,
"env/all/time/reward_compute": 6.007030606269836e-07,
"env/all/time/reward_compute/min": 2.2351741790771484e-07,
"env/all/time/reward_compute/max": 1.385807991027832e-06,
"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": 0.00563928484916687,
"advantage/min": -1.0,
"advantage/max": 0.8031213283538818,
"time/assemble_training_data": 10.164772510528564,
"time/kl_vs_base": 147.0721833705902,
"kl_policy_base": 0.0006728654261678457,
"time/train": 1175.305495262146,
"time/save_checkpoint": 22.228652000427246,
"time/total": 4755.478853225708
}
[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