Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step45
ttt-discover-v3-qwen3-8b-ac1-step45 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 44 (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 44 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 44,
"progress/batch": 44,
"optim/lr": 4e-05,
"progress/done_frac": 0.9,
"puct/buffer_size": 712,
"puct/sampled_size": 8,
"puct/T": 22528,
"puct/scale_last": 251.17629545984107,
"puct/buffer_value/mean": -1.8768583868089777,
"puct/buffer_value/std": 9.406278147232015,
"puct/buffer_value/min": -252.6826286848707,
"puct/buffer_value/max": -1.5063332250296213,
"puct/buffer_timestep/mean": 21.247191011235955,
"puct/buffer_timestep/std": 12.847664850453647,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 43.0,
"puct/buffer_construction_len/mean": 1061.1390449438202,
"puct/buffer_construction_len/std": 503.847649573867,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5063332250296235,
"puct/sampled_value/std": 2.0137621733714643e-15,
"puct/sampled_value/min": -1.506333225029626,
"puct/sampled_value/max": -1.5063332250296213,
"puct/sampled_timestep/mean": 42.125,
"puct/sampled_timestep/std": 0.9270248108869579,
"puct/sampled_timestep/min": 41.0,
"puct/sampled_timestep/max": 43.0,
"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": 3373.0002982616425,
"env/all/ac_tokens_per_turn": 8850.939453125,
"env/all/ob_tokens_per_turn": 2765.875,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4531681,
"env/all/total_ob_tokens": 1416128,
"env/all/time/sampling_mean": 310.1323555414565,
"env/all/time/sampling_max": 412.76338291168213,
"env/all/time/env_step_mean": 1128.743399572093,
"env/all/time/env_step_max": 2949.665919780731,
"env/all/reward/mean": 0.6012869793199456,
"env/all/reward/max": 0.6638637299802627,
"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": 0.6012869793199456,
"env/all/correctness": 0.916015625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.529292217848696,
"env/all/raw_score/min": 1.5063332250296213,
"env/all/raw_score/max": 3.0030030030030033,
"env/all/initial_raw_score": -1.5063332250296233,
"env/all/initial_raw_score/min": -1.506333225029626,
"env/all/initial_raw_score/max": -1.5063332250296213,
"env/all/msg": "Success; raw_score=1.5063332250296213",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy.optimize import linprog\nfrom scipy.optimize import minimize\n\ndef propose_candidate(seed=42, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score \n using an evolved approach with dynamic constraint selection, adaptive perturbation,\n and enhanced solver configurations for deeper exploration.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + 1000 - 30 # Run for up to 1000 seconds with 30s buffer\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 normalized to sqrt(2n)\n n = 1000\n initial_sequence = np.random.uniform(low=0.1, high=1.0, size=n)\n initial_sequence = initial_sequence / np.sum(initial_sequence) * np.sqrt(2 * n)\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 exploration and learning\n k_values = [50, 100, 200] # Dynamic number of constraints\n perturbation_steps = 200\n perturbation_scale = 0.05 # Increased perturbation scale for better exploration\n alpha_search_iterations = 200 # More precise alpha tuning\n adaptive_learning_rate = True # Enable learning rate adjustment\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 adaptively 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 A = []\n for j in top_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(top_indices), max_b * 1.01) # Slight relaxation of the constraint\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' with tighter tolerances\n res = linprog(c=c, A_ub=A, b_ub=b, bounds=bounds, method='highs', \n options={\"feastol\": 1e-9, \"inf_tol\": 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 # Normalize g_0 to match the theoretical normalization sum\n g_0_normalized = g_0 / sum_g * np.sqrt(2 * n)\n\n # Use ternary search to find optimal alpha with dynamic learning rate\n def evaluate_alpha(alpha):\n new_sequence = (1 - alpha) * curr_sequence + alpha * g_0_normalized\n conv_new = np.convolve(new_sequence, new_sequence)\n new_max_b = np.max(conv_new)\n return 2 * n * new_max_b / (np.sum(new_sequence) ** 2)\n\n low, high = 0.0, 1.0\n for _ in range(alpha_search_iterations):\n m1 = low + (high - low)/3\n m2 = high - (high - low)/3\n if evaluate_alpha(m1) < evaluate_alpha(m2):\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 new_sequence = new_sequence / np.sum(new_sequence) * np.sqrt(2 * n)\n curr_sequence = new_sequence.copy()\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 # 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:\n print(\"No improvement for 200 iterations, stopping early\")\n break\n\n # Introduce more aggressive random perturbations to escape local minima\n if iteration_count % 50 == 0:\n perturb = np.random.normal(0, perturbation_scale, size=n)\n curr_sequence += perturb\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 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": 310.1323555414565,
"env/all/time/policy/min": 146.13696908950806,
"env/all/time/policy/max": 412.76338291168213,
"env/all/time/env_step": 1128.743399572093,
"env/all/time/env_step/min": 0.5956573486328125,
"env/all/time/env_step/max": 2949.665919780731,
"env/all/time/reward_compute": 4.4284388422966003e-07,
"env/all/time/reward_compute/min": 2.2724270820617676e-07,
"env/all/time/reward_compute/max": 8.270144462585449e-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": 0.002809010911732912,
"advantage/min": -1.0,
"advantage/max": 0.5365853309631348,
"time/assemble_training_data": 5.864724159240723,
"time/kl_vs_base": 94.12783455848694,
"kl_policy_base": 0.0006922274478711188,
"time/train": 559.5183000564575,
"time/save_checkpoint": 19.390414714813232,
"time/total": 4053.9275217056274
}
[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
[2026-07-12T04:58:24+00:00] job=1827204 node=node-4 ngpu=6 ntrain=2 replicas=4 flash_attn=yes