Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step14
ttt-discover-v3-qwen3-8b-ac1-step14 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 13 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 10, 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 13 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 13,
"progress/batch": 13,
"optim/lr": 4e-05,
"progress/done_frac": 0.28,
"puct/buffer_size": 216,
"puct/sampled_size": 8,
"puct/T": 6656,
"puct/scale_last": 0.3992066724615333,
"puct/buffer_value/mean": -1.5364794017824512,
"puct/buffer_value/std": 0.10452479395602289,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5064163683325371,
"puct/buffer_timestep/mean": 5.7407407407407405,
"puct/buffer_timestep/std": 3.9024455067027692,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 12.0,
"puct/buffer_construction_len/mean": 1201.5324074074074,
"puct/buffer_construction_len/std": 899.1729836384741,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5064892081147099,
"puct/sampled_value/std": 4.3687171978636475e-05,
"puct/sampled_value/min": -1.5065379412051672,
"puct/sampled_value/max": -1.5064163683325371,
"puct/sampled_timestep/mean": 12.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 12.0,
"puct/sampled_timestep/max": 12.0,
"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": 5323.630368471146,
"env/all/ac_tokens_per_turn": 8744.130859375,
"env/all/ob_tokens_per_turn": 3423.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4476995,
"env/all/total_ob_tokens": 1752576,
"env/all/time/sampling_mean": 325.5572414738126,
"env/all/time/sampling_max": 419.82508182525635,
"env/all/time/env_step_mean": 2643.2298463825136,
"env/all/time/env_step_max": 4908.383753538132,
"env/all/reward/mean": 0.5778987748679295,
"env/all/reward/max": 0.6638403549322465,
"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": 0.5778987748679295,
"env/all/correctness": 0.91015625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 2.201915966000778,
"env/all/raw_score/min": 1.506386265811845,
"env/all/raw_score/max": 288.51294228853516,
"env/all/initial_raw_score": -1.5064892081147097,
"env/all/initial_raw_score/min": -1.5065379412051672,
"env/all/initial_raw_score/max": -1.5064163683325371,
"env/all/msg": "Success; raw_score=1.5064076685452588",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy import optimize\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score \n using enhanced LP-based directions, adaptive perturbations, and optimized LP solving.\n The algorithm prioritizes targeted perturbations and improves LP solving to explore better sequences.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Initialize using the best known sequence or a structured uniform + Gaussian blend\n n = 1000\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 # Uniform and Gaussian-initialized sequence, clipped and normalized\n uniform_seq = np.ones(n) * np.sqrt(2 * n) / n\n gaussian_seq = np.random.normal(0.5, 0.15, n)\n gaussian_seq = np.maximum(0.0, np.clip(gaussian_seq, 0, 1000.0))\n gaussian_seq *= np.sqrt(2 * n) / np.sum(gaussian_seq)\n best_sequence = [float(x) for x in gaussian_seq * 0.7 + uniform_seq * 0.3]\n \n curr_sequence = [float(x) for x in best_sequence]\n best_score = evaluate_sequence(curr_sequence)\n\n # Parameters for optimization\n perturb_rate = 0.05 # Reduced perturbation rate for controlled changes\n target_perturb = 0.2 # Perturbation fraction (not directly used)\n min_perturb = 0.01\n max_perturb_attempts = 3 # Limit number of perturbation tries\n perturbation_radius = 0.1 # Radius for perturbing elements\n\n while time.time() < deadline:\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 < 0.01:\n raise ValueError(\"Sum too small\")\n\n # Normalize for LP\n normalized_seq = [x / sum_a * np.sqrt(2 * n) for x in curr_sequence]\n rhs = max_b\n\n # Solve LP to maximize sum(b) under convolution constraint\n g_fun = solve_convolution_lp(normalized_seq, rhs)\n if g_fun is not None and np.sum(g_fun) > 0.0:\n sum_g = np.sum(g_fun)\n normalized_g = [x / sum_g * np.sqrt(2 * n) for x in g_fun]\n \n # Evaluate the LP solution directly\n new_score = evaluate_sequence(normalized_g)\n if new_score < best_score:\n curr_sequence = normalized_g\n curr_score = new_score\n else:\n # Proceed with blending (limited to small steps)\n def objective(t):\n new_seq = [(1 - t) * x + t * y for x, y in zip(curr_sequence, normalized_g)]\n return evaluate_sequence(new_seq)\n \n result = optimize.minimize_scalar(objective, bounds=(0.0, 1.0), method='bounded', tol=1e-5)\n best_t = result.x\n new_seq = [(1 - best_t) * x + best_t * y for x, y in zip(curr_sequence, normalized_g)]\n curr_sequence = new_seq\n else:\n # Fallback to targeted perturbation based on max_b contributions\n max_index = np.argmax(conv)\n contributions = [abs(curr_sequence[i] * curr_sequence[max_index - i]) for i in range(n)]\n contribution_indices = np.argsort(contributions)[::-1][:min(5, n)] # Top contributions\n\n # Try perturbing these indices\n for _ in range(min(max_perturb_attempts, 5)):\n if len(contribution_indices) == 0:\n idx = np.random.randint(0, n)\n curr_sequence[idx] = max(0.0, curr_sequence[idx] - perturb_rate * curr_sequence[idx])\n else:\n idx = contribution_indices[np.random.randint(len(contribution_indices))]\n curr_sequence[idx] = max(0.0, curr_sequence[idx] - perturb_rate * curr_sequence[idx])\n curr_sequence[idx] = min(1000.0, curr_sequence[idx])\n\n except Exception as e:\n print(f\"Error during optimization: {e}. Skipping update.\")\n continue\n\n try:\n curr_score = evaluate_sequence(curr_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = curr_sequence.copy()\n print(f\"New best score: {best_score}\")\n except Exception:\n print(\"Evaluation error, skipping update\")\n\n return [float(max(0.0, x)) for x in best_sequence]\n\ndef solve_convolution_lp(f_sequence, rhs):\n \"\"\"Solves LP to maximize sum(b) s.t. conv(f, b) <= rhs, b >= 0 with increased time limit.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n c = -np.ones(n) # Maximize sum(b) via -1 coefficients\n a_ub = []\n b_ub = []\n\n for k in range(2 * n - 1):\n row = np.zeros(n)\n for i in range(n):\n j = k - i\n if 0 <= j < n:\n row[j] = f_sequence[i]\n a_ub.append(row)\n b_ub.append(rhs)\n\n # Non-negativity constraints\n a_ub_nonneg = -np.eye(n)\n b_ub_nonneg = np.zeros(n)\n a_ub = np.vstack([a_ub, a_ub_nonneg])\n b_ub = np.hstack([b_ub, b_ub_nonneg])\n\n try:\n result = optimize.linprog(\n c,\n A_ub=a_ub,\n b_ub=b_ub,\n bounds=(0.0, 1000.0),\n method='highs',\n options={\n \"time_limit\": 120.0, # Increased time limit for LP\n \"disp\": False,\n },\n )\n if result.success:\n return result.x\n except:\n pass\n return None\n```",
"env/all/time/policy": 325.5572414738126,
"env/all/time/policy/min": 192.34166288375854,
"env/all/time/policy/max": 419.82508182525635,
"env/all/time/env_step": 2643.2298463825136,
"env/all/time/env_step/min": 0.008611440658569336,
"env/all/time/env_step/max": 4908.383753538132,
"env/all/time/reward_compute": 3.4831464290618896e-07,
"env/all/time/reward_compute/min": 2.1606683731079102e-07,
"env/all/time/reward_compute/max": 6.593763828277588e-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": 0.012577855959534645,
"advantage/min": -1.0,
"advantage/max": 2.8645567893981934,
"time/assemble_training_data": 5.917264461517334,
"time/kl_vs_base": 93.42253684997559,
"kl_policy_base": 0.000781459326390177,
"time/train": 585.2960252761841,
"time/save_checkpoint": 34.57281517982483,
"time/total": 6044.471734762192
}
[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