Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step47
ttt-discover-v3-qwen3-8b-ac1-step47 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 46 (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 46 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 46,
"progress/batch": 46,
"optim/lr": 4e-05,
"progress/done_frac": 0.94,
"puct/buffer_size": 742,
"puct/sampled_size": 8,
"puct/T": 23552,
"puct/scale_last": 251.17629545984107,
"puct/buffer_value/mean": -1.868961371789111,
"puct/buffer_value/std": 9.214452099497821,
"puct/buffer_value/min": -252.6826286848707,
"puct/buffer_value/max": -1.5063332250296213,
"puct/buffer_timestep/mean": 22.18598382749326,
"puct/buffer_timestep/std": 13.39088549456605,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 45.0,
"puct/buffer_construction_len/mean": 1058.667115902965,
"puct/buffer_construction_len/std": 493.70385284041663,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5063332250296235,
"puct/sampled_value/std": 1.6203171603697468e-15,
"puct/sampled_value/min": -1.506333225029626,
"puct/sampled_value/max": -1.5063332250296213,
"puct/sampled_timestep/mean": 44.5,
"puct/sampled_timestep/std": 1.3228756555322954,
"puct/sampled_timestep/min": 41.0,
"puct/sampled_timestep/max": 45.0,
"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": 2849.7054624557495,
"env/all/ac_tokens_per_turn": 8632.294921875,
"env/all/ob_tokens_per_turn": 2872.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4419735,
"env/all/total_ob_tokens": 1470784,
"env/all/time/sampling_mean": 295.840510701295,
"env/all/time/sampling_max": 406.6435103416443,
"env/all/time/env_step_mean": 972.2126662335359,
"env/all/time/env_step_max": 2453.2290074825287,
"env/all/reward/mean": 0.5745143293411283,
"env/all/reward/max": 0.6638637299802628,
"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": 0.5745143293411283,
"env/all/correctness": 0.8828125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.7797061620359762,
"env/all/raw_score/min": 1.506333225029621,
"env/all/raw_score/max": 93.53772506043926,
"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 by\n using a more effective constraint selection in the linear program and adjusting\n the alpha search.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + 1000\n\n # Initialize with a random sequence or load from a previous best\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 # Start with a random distribution\n n = 1000\n initial_sequence = np.random.randn(n) * 2.0 # Spread values more\n initial_sequence = np.maximum(0.0, initial_sequence)\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 perturbation_scale = 0.1 # Start with larger perturbations\n alpha_search_iterations = 100\n\n # Golden section search parameters\n def golden_section_search(f, low, high, max_iter=100):\n \"\"\"Minimize unimodal function f with golden section search.\"\"\"\n phi = (1 + np.sqrt(5)) / 2\n c = 1 - 1/phi\n d = 1/phi\n x1 = high - d * (high - low)\n x2 = low + d * (high - low)\n for _ in range(max_iter):\n f1 = f(x1)\n f2 = f(x2)\n if f1 < f2:\n high = x2\n x2 = x1\n x1 = high - d * (high - low)\n else:\n low = x1\n x1 = x2\n x2 = low + d * (high - low)\n return (low + high) / 2\n\n # Track the maximum index for tighter constraints\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 # Dynamic constraint selection\n # Select constraints where current convolution is above a threshold\n threshold = max_b * 0.95\n high_indices = [i for i, val in enumerate(conv) if val >= threshold]\n A = []\n for j in high_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(high_indices), max_b) # Use the current max_b as 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' 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 # Evaluate function for golden section search\n def evaluate_alpha(alpha):\n new_sequence = (1 - alpha) * curr_sequence + alpha * g_0\n new_sequence = np.clip(new_sequence, 0.0, 1000.0)\n conv_new = np.convolve(new_sequence, new_sequence)\n new_max_b = np.max(conv_new)\n sum_new = np.sum(new_sequence)\n return 2 * n * new_max_b / (sum_new ** 2) # Simplified for line search\n\n # Golden section search for optimal alpha\n low, high = 0.0, 1.0\n alpha_opt = golden_section_search(evaluate_alpha, low, high, max_iter=100)\n new_sequence = (1 - alpha_opt) * curr_sequence + alpha_opt * g_0\n\n # Clip values and normalize\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\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 # Random perturbation if LP fails\n perturb = np.random.normal(0, perturbation_scale, size=n)\n curr_sequence = np.clip(curr_sequence + perturb, 0.0, 1000.0)\n curr_sequence = curr_sequence / np.sum(curr_sequence) * np.sqrt(2 * n)\n\n # Gradually reduce perturbation scale\n if iteration_count % 100 == 0:\n perturbation_scale *= 0.98 # Reduce perturbations over time\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 except Exception as e:\n print(f\"Error during optimization: {e}. Skipping update.\")\n continue\n\n # Final safety check and return\n final_sequence = np.maximum(0.0, best_sequence).tolist()\n return [float(x) for x in final_sequence]\n```",
"env/all/time/policy": 295.840510701295,
"env/all/time/policy/min": 124.60144901275635,
"env/all/time/policy/max": 406.6435103416443,
"env/all/time/env_step": 972.2126662335359,
"env/all/time/env_step/min": 0.741131067276001,
"env/all/time/env_step/max": 2453.2290074825287,
"env/all/time/reward_compute": 3.180466592311859e-07,
"env/all/time/reward_compute/min": 2.2351741790771484e-07,
"env/all/time/reward_compute/max": 7.37607479095459e-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": 0.003980787470936775,
"advantage/min": -1.0,
"advantage/max": 0.5365853309631348,
"time/assemble_training_data": 9.69556713104248,
"time/kl_vs_base": 95.81131505966187,
"kl_policy_base": 0.0006743308040313423,
"time/train": 552.0257048606873,
"time/save_checkpoint": 17.734302759170532,
"time/total": 3527.3850104808807
}
[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