Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step16
ttt-discover-v3-qwen3-8b-ac1-s2-step16 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 15 (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 15 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 15,
"progress/batch": 15,
"optim/lr": 4e-05,
"progress/done_frac": 0.32,
"puct/buffer_size": 248,
"puct/sampled_size": 8,
"puct/T": 7680,
"puct/scale_last": 0.5736105308137764,
"puct/buffer_value/mean": -1.5480956422577223,
"puct/buffer_value/std": 0.12211052268018677,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5061769375344798,
"puct/buffer_timestep/mean": 6.741935483870968,
"puct/buffer_timestep/std": 4.479110956828331,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 14.0,
"puct/buffer_construction_len/mean": 1241.8951612903227,
"puct/buffer_construction_len/std": 1031.4614109285192,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5062453345492934,
"puct/sampled_value/std": 5.858758804767605e-05,
"puct/sampled_value/min": -1.5063036919506334,
"puct/sampled_value/max": -1.5061769375344798,
"puct/sampled_timestep/mean": 13.875,
"puct/sampled_timestep/std": 0.33071891388307384,
"puct/sampled_timestep/min": 13.0,
"puct/sampled_timestep/max": 14.0,
"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": 5329.223585605621,
"env/all/ac_tokens_per_turn": 8129.326171875,
"env/all/ob_tokens_per_turn": 2894.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4162215,
"env/all/total_ob_tokens": 1481728,
"env/all/time/sampling_mean": 274.327512959484,
"env/all/time/sampling_max": 371.281352519989,
"env/all/time/env_step_mean": 2516.2522430950776,
"env/all/time/env_step_max": 4985.367208003998,
"env/all/reward/mean": 0.5355480890444866,
"env/all/reward/max": 0.6640553147937269,
"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": 0.5355480890444866,
"env/all/correctness": 0.830078125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 5.693128208113935,
"env/all/raw_score/min": 1.5058986368779734,
"env/all/raw_score/max": 847.3597332061167,
"env/all/initial_raw_score": -1.5062453345492937,
"env/all/initial_raw_score/min": -1.5063036919506334,
"env/all/initial_raw_score/max": -1.5061769375344798,
"env/all/msg": "Success; raw_score=1.5061426845636772",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nimport cvxpy as cp\nimport copy\nimport random\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Start from known best construction if it exists\n prev = globals().get(\"height_sequence_1\", None)\n if prev is not None and isinstance(prev, (list, np.ndarray)) and len(prev) > 0:\n best_sequence = list(np.asarray(prev, dtype=float))\n else:\n n = 1000\n # Generate multiple initial sequences for diverse starting points\n initial_sequences = []\n for _ in range(3):\n if np.random.rand() < 0.25:\n # Uniform distribution with higher density\n base = np.random.uniform(0.01, 0.05, size=n)\n elif np.random.rand() < 0.5:\n # Uniform distribution with lower density\n base = np.random.uniform(0.0001, 0.001, size=n)\n elif np.random.rand() < 0.75:\n # Exponential distribution for more variance\n base = np.random.exponential(scale=0.01, size=n)\n else:\n # Poisson distribution for sparse spikes\n base = np.random.poisson(lam=0.1, size=n)\n sum_base = np.sum(base)\n scale_factor = 0.1 / sum_base\n initial_sequences.append([max(0.0, x * scale_factor) for x in base])\n best_sequence = initial_sequences[0].copy()\n for seq in initial_sequences[1:]:\n curr_score = evaluate_sequence(seq)\n best_score = evaluate_sequence(best_sequence)\n if curr_score < best_score:\n best_sequence = seq.copy()\n \n current_sequence = best_sequence.copy()\n best_score = float('inf')\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using dynamic constraint selection and adaptive line search.\"\"\"\n n = len(sequence)\n if n == 0:\n return None\n\n sum_sequence = np.sum(sequence)\n if sum_sequence <= 0.0:\n return None\n\n conv = np.convolve(sequence, sequence)\n max_b_val = np.max(conv)\n # Dynamic tight position selection with adaptive sensitivity\n tolerance = max_b_val * 0.15 # Increased tolerance to include more tight positions\n tight_positions = np.where(conv >= max_b_val - tolerance)[0]\n if len(tight_positions) < int(len(conv)*0.5):\n # Fallback to top 50% of positions\n tight_positions = np.argsort(conv)[::-1][:int(len(conv)*0.5)]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n\n # Solve LP with extended time limit\n g_fun = solve_convolution_lp(sequence, max_b_val, tight_positions)\n if g_fun is None:\n return None\n\n sum_g = np.sum(g_fun)\n if sum_g <= 0.0:\n return None\n\n # Improved line search with golden-section search for optimal t\n def golden_section_search(f, a, b):\n phi = (np.sqrt(5) - 1) / 2\n c = b - phi * (b - a)\n d = a + phi * (b - a)\n while abs(b - a) > 1e-8:\n if f(c) < f(d):\n b = d\n else:\n a = c\n c = b - phi * (b - a)\n d = a + phi * (b - a)\n return (a + b) / 2\n\n def evaluate_t(t):\n new_seq = [(1 - t)*x + t*y for x, y in zip(sequence, g_fun)]\n try:\n return evaluate_sequence(new_seq)\n except:\n return float('inf')\n\n # Initial bounds for t\n low, high = 0.01, 0.99\n t_opt = golden_section_search(evaluate_t, low, high)\n new_seq = [(1 - t_opt)*x + t_opt*y for x, y in zip(sequence, g_fun)]\n return new_seq\n\n def solve_convolution_lp(f_sequence, rhs, tight_positions):\n \"\"\"Solves LP with adaptive constraint selection.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n # Variables\n g = cp.Variable(n, nonneg=True)\n # Objective: maximize sum(g_j), equivalent to minimize -sum(g_j)\n objective = cp.Minimize(-cp.sum(g))\n\n # Constraints\n constraints = []\n for k in tight_positions:\n # Coefficients for the constraint: sum_{j=0}^{n-1} f[i] * g[j] <= rhs for appropriate i\n coeff = np.zeros(n)\n for j in range(n):\n i = k - j\n if 0 <= i < n:\n coeff[j] = f_sequence[i]\n constraints.append(cp.sum(coeff * g) <= rhs)\n\n # Solve the LP\n problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.GLPK, verbose=False)\n if problem.status == cp.OPTIMAL:\n return g.value\n except:\n return None\n return None\n\n def perturb_sequence(seq):\n \"\"\"Perturbation with diverse strategies to explore local minima.\"\"\"\n conv = np.convolve(seq, seq)\n # Select top 20 indices with highest convolution values\n top_indices = np.argsort(conv)[-20:]\n # Randomly choose 20 distinct indices from the top 20 for perturbation\n indices = np.random.choice(top_indices, size=20, replace=False)\n indices = [min(idx, len(seq)-1) for idx in indices]\n for idx in indices:\n # Larger random variation with exponential scaling\n scale = 0.1 * np.std(seq)\n new_val = max(0.0, seq[idx] + np.random.exponential(scale))\n seq = [new_val if i == idx else seq[i] for i in range(len(seq))]\n return seq\n\n def random_walk(seq):\n \"\"\"Adds small random perturbations to all elements.\"\"\"\n scale = 0.1 * np.std(seq)\n new_seq = [max(0.0, x + np.random.normal(0, scale)) for x in seq]\n return new_seq\n\n while time.time() < deadline:\n # Try LP-based move\n h_function = get_good_direction_to_move_into(current_sequence)\n if h_function is None:\n current_sequence = perturb_sequence(current_sequence)\n else:\n current_sequence = h_function\n\n # Occasionally try random walk for exploration\n if np.random.rand() < 0.15:\n current_sequence = random_walk(current_sequence)\n\n try:\n curr_score = evaluate_sequence(current_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = current_sequence.copy()\n print(f\"New best: {best_score}\")\n except Exception:\n pass\n\n return [float(max(0.0, x)) for x in best_sequence]\n```",
"env/all/time/policy": 274.327512959484,
"env/all/time/policy/min": 128.97767758369446,
"env/all/time/policy/max": 371.281352519989,
"env/all/time/env_step": 2516.2522430950776,
"env/all/time/env_step/min": 0.009357452392578125,
"env/all/time/env_step/max": 4985.367208003998,
"env/all/time/reward_compute": 5.471520125865936e-07,
"env/all/time/reward_compute/min": 1.862645149230957e-07,
"env/all/time/reward_compute/max": 2.4549663066864014e-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": 0.025383424013853073,
"advantage/min": -1.0,
"advantage/max": 19.923965454101562,
"time/assemble_training_data": 5.411512613296509,
"time/kl_vs_base": 88.35421586036682,
"kl_policy_base": 0.0008697897428646684,
"time/train": 521.8526556491852,
"time/save_checkpoint": 8.518989324569702,
"time/total": 5955.384048700333
}
[2026-07-09T06:24:18+00:00] job=1812630 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T09:14:17+00:00] job=1813129 node=node-28 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:46:36+00:00] job=1813130 node=node-22 ngpu=6 ntrain=2 replicas=4 flash_attn=yes