Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step5
ttt-discover-v3-qwen3-8b-ac1-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": 0.39798783792233383,
"puct/buffer_value/mean": -1.5940859088316233,
"puct/buffer_value/std": 0.1667219783667827,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5076352028717366,
"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": 1604.5972222222222,
"puct/buffer_construction_len/std": 1477.1067314528746,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5097954392238593,
"puct/sampled_value/std": 0.0008910805157816385,
"puct/sampled_value/min": -1.5103659125339968,
"puct/sampled_value/max": -1.5076352028717366,
"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": 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": 5242.339818477631,
"env/all/ac_tokens_per_turn": 8597.1484375,
"env/all/ob_tokens_per_turn": 2853.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4401740,
"env/all/total_ob_tokens": 1460800,
"env/all/time/sampling_mean": 300.7665830575861,
"env/all/time/sampling_max": 384.383996963501,
"env/all/time/env_step_mean": 2131.711717688013,
"env/all/time/env_step_max": 4866.9879739284515,
"env/all/reward/mean": 0.48496363019957045,
"env/all/reward/max": 0.6634002281528624,
"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": 0.48496363019957045,
"env/all/correctness": 0.775390625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 6.653450067685748,
"env/all/raw_score/min": 1.5073856639307264,
"env/all/raw_score/max": 1991.149512529431,
"env/all/initial_raw_score": -1.5097954392238593,
"env/all/initial_raw_score/min": -1.5103659125339968,
"env/all/initial_raw_score/max": -1.5076352028717366,
"env/all/msg": "Success; raw_score=1.5076352028717366",
"env/all/parsed_code": "```python\nimport numpy as np\nimport time\nimport random\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Proposes an optimal sequence of non-negative numbers that minimizes the evaluation score\n using gradient descent with adaptive learning and perturbation to avoid local minima.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n \n # Initialize with the best sequence if available\n best_sequence = None\n if 'height_sequence_1' in globals() and isinstance(height_sequence_1, (list, np.ndarray)):\n best_sequence = list(height_sequence_1)\n else:\n n = 1000\n gaussian_seq = np.exp(-np.arange(n)**2 / (n**2)) / np.sum(np.exp(-np.arange(n)**2 / (n**2)))\n best_sequence = gaussian_seq.tolist()\n \n curr_sequence = [float(max(0.0, x)) for x in best_sequence]\n best_score = evaluate_sequence(curr_sequence)\n \n # Learning rate and decay factors\n learning_rate = 0.5\n decay = 0.99\n grad_threshold = 1e-5\n max_iterations = 10000\n \n # Gradient computation\n def compute_gradient(seq):\n n = len(seq)\n conv = np.convolve(seq, seq)\n max_B = np.max(conv)\n if max_B == 0:\n return np.zeros(n)\n k = np.argmax(conv)\n grads = np.zeros(n)\n for i in range(n):\n j = k - i\n if 0 <= j < n:\n grads[i] = seq[j]\n S = np.sum(seq)\n gradient = (2 * n * (grads * S - 2 * max_B)) / (S**3)\n return gradient\n \n iteration = 0\n while time.time() < deadline and iteration < max_iterations:\n try:\n gradient = compute_gradient(curr_sequence)\n gradient_norm = np.linalg.norm(gradient)\n \n # Update sequence\n new_seq = curr_sequence - learning_rate * gradient\n new_seq = np.clip(new_seq, 0.0, 1000.0)\n \n # Check sum and scale if necessary\n sum_new = np.sum(new_seq)\n if sum_new < 0.01:\n scale_factor = 0.01 / sum_new\n new_seq = np.clip(new_seq * scale_factor, 0.0, 1000.0)\n \n # Update current sequence\n curr_sequence = new_seq.copy()\n \n # Evaluate new sequence\n new_score = evaluate_sequence(curr_sequence)\n if new_score < best_score:\n best_score = new_score\n best_sequence = curr_sequence.copy()\n print(f\"New best score: {best_score}\")\n \n # Check gradient for stagnation\n if gradient_norm < grad_threshold:\n print(\"Gradient near zero, perturbing sequence.\")\n perturbation = np.random.randn(len(curr_sequence)) * 0.5\n curr_sequence = np.clip(curr_sequence + perturbation, 0.0, 1000.0)\n sum_curr = np.sum(curr_sequence)\n if sum_curr < 0.01:\n scale_factor = 0.01 / sum_curr\n curr_sequence = np.clip(curr_sequence * scale_factor, 0.0, 1000.0)\n \n # Decay learning rate\n learning_rate *= decay\n iteration += 1\n \n except Exception as e:\n print(f\"Error during iteration: {e}, skipping update\")\n continue\n \n return [float(max(0.0, x)) for x in best_sequence]\n```",
"env/all/time/policy": 300.7665830575861,
"env/all/time/policy/min": 35.27770113945007,
"env/all/time/policy/max": 384.383996963501,
"env/all/time/env_step": 2131.711717688013,
"env/all/time/env_step/min": 0.007437705993652344,
"env/all/time/env_step/max": 4866.9879739284515,
"env/all/time/reward_compute": 5.699694156646729e-07,
"env/all/time/reward_compute/min": 2.1979212760925293e-07,
"env/all/time/reward_compute/max": 2.738088369369507e-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": 0.02645636349916458,
"advantage/min": -1.0,
"advantage/max": 15.544912338256836,
"time/assemble_training_data": 9.70430040359497,
"time/kl_vs_base": 90.25493812561035,
"kl_policy_base": 0.000596971542108804,
"time/train": 546.0055530071259,
"time/save_checkpoint": 34.828978061676025,
"time/total": 5924.9325494766235
}
[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