Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step28
ttt-discover-v3-qwen3-8b-ac1-step28 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 27 (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 27 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 27,
"progress/batch": 27,
"optim/lr": 4e-05,
"progress/done_frac": 0.56,
"puct/buffer_size": 440,
"puct/sampled_size": 8,
"puct/T": 13824,
"puct/scale_last": 0.4936516100901467,
"puct/buffer_value/mean": -1.5232084185587171,
"puct/buffer_value/std": 0.07914682911494018,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5063483899098538,
"puct/buffer_timestep/mean": 12.745454545454546,
"puct/buffer_timestep/std": 7.941188786248755,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 26.0,
"puct/buffer_construction_len/mean": 1098.9340909090909,
"puct/buffer_construction_len/std": 638.0097090466647,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5063602554243312,
"puct/sampled_value/std": 1.04402647537122e-11,
"puct/sampled_value/min": -1.5063602554370648,
"puct/sampled_value/max": -1.5063602553985826,
"puct/sampled_timestep/mean": 26.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 26.0,
"puct/sampled_timestep/max": 26.0,
"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": 5449.050518512726,
"env/all/ac_tokens_per_turn": 8441.419921875,
"env/all/ob_tokens_per_turn": 3014.5,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4322007,
"env/all/total_ob_tokens": 1543424,
"env/all/time/sampling_mean": 297.02721296297386,
"env/all/time/sampling_max": 381.9443619251251,
"env/all/time/env_step_mean": 2714.6749688144773,
"env/all/time/env_step_max": 5066.00408411026,
"env/all/reward/mean": 0.6161190137194248,
"env/all/reward/max": 0.6638518175094403,
"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": 0.6161190137194248,
"env/all/correctness": 0.958984375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.8341566804038736,
"env/all/raw_score/min": 1.5063602553852484,
"env/all/raw_score/max": 133.33333333333331,
"env/all/initial_raw_score": -1.5063602554243312,
"env/all/initial_raw_score/min": -1.5063602554370648,
"env/all/initial_raw_score/max": -1.5063602553985826,
"env/all/msg": "Success; raw_score=1.5063602553982236",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy.optimize import linprog\nimport random\nimport copy\nimport math\n\ndef propose_candidate(seed=42, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score \n using an optimized initialization, constraint selection, and line search\n strategy with enhanced exploratory and adaptability features.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + 1000 - 10 # Run for up to 1000 seconds\n\n n = 1000\n # Start with a high-quality initial sequence if available, otherwise a uniform distribution\n best_sequence = None\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 uniform distribution\n best_sequence = np.random.rand(n) * 0.01\n # Ensure non-zero sum\n best_sequence = np.clip(best_sequence, 0.001, 1.0)\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 # Use golden-section search for alpha optimization\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 # Build constraints for LP with active indices\n # Select active indices by considering top 10% of conv values\n conv_sorted = np.sort(conv)\n threshold = conv_sorted[-int(len(conv) * 0.1)]\n active_indices = np.where(conv >= threshold)[0]\n if len(active_indices) == 0:\n raise ValueError(\"No active constraints found\")\n\n A = []\n for j in active_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(active_indices), max_b)\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)] # Use higher upper bound\n\n # Solve the LP using the 'highs' method for better 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 # Perform line search to find optimal scaling factor\n def score_func(scale):\n scaled_g = g_0 * scale\n new_sequence = scaled_g\n return evaluate_sequence(new_sequence.tolist())\n \n a, b = 0.0, 20.0 # Scale can be up to 20\n iterations = 200\n for _ in range(iterations):\n c = a + (b - a)/2\n d = a + (b - a)*0.381966\n f_c = score_func(c)\n f_d = score_func(d)\n if f_d < f_c:\n b = c\n else:\n a = d\n best_scale = (a + b)/2\n g_0_normalized = g_0 * best_scale\n g_0_normalized = np.clip(g_0_normalized, 0.0, 1000.0)\n\n # Define objective function for golden-section search\n def score_func(alpha):\n new_sequence = (1 - alpha) * curr_sequence + alpha * g_0_normalized\n # Clip to ensure non-negative and within bounds\n new_sequence = np.clip(new_sequence, 0.0, 1000.0)\n return evaluate_sequence(new_sequence.tolist())\n\n # Use golden-section search to find the optimal alpha between 0 and 1\n a, b = 0.0, 1.0\n iterations = 200\n for _ in range(iterations):\n c = a + (b - a)/2\n d = a + (b - a)*0.381966\n f_c = score_func(c)\n f_d = score_func(d)\n if f_d < f_c:\n b = c\n else:\n a = d\n best_alpha = (a + b) / 2\n \n # Apply the best alpha found\n curr_sequence = (1 - best_alpha) * curr_sequence + best_alpha * g_0_normalized\n curr_sequence = np.clip(curr_sequence, 0.0, 1000.0)\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 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": 297.02721296297386,
"env/all/time/policy/min": 115.6459550857544,
"env/all/time/policy/max": 381.9443619251251,
"env/all/time/env_step": 2714.6749688144773,
"env/all/time/env_step/min": 0.9851157665252686,
"env/all/time/env_step/max": 5066.00408411026,
"env/all/time/reward_compute": 3.203749656677246e-07,
"env/all/time/reward_compute/min": 2.2724270820617676e-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": 0.007165214978158474,
"advantage/min": -1.0,
"advantage/max": 0.8529410362243652,
"time/assemble_training_data": 9.531835079193115,
"time/kl_vs_base": 90.2046582698822,
"kl_policy_base": 0.000761279312428087,
"time/train": 544.6164028644562,
"time/save_checkpoint": 34.39506530761719,
"time/total": 6130.106067419052
}
[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