Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step23
ttt-discover-v3-qwen3-8b-ac1-s2-step23 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 22 (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 22 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 22,
"progress/batch": 22,
"optim/lr": 4e-05,
"progress/done_frac": 0.46,
"puct/buffer_size": 360,
"puct/sampled_size": 8,
"puct/T": 11264,
"puct/scale_last": 0.574125750004278,
"puct/buffer_value/mean": -1.5349553383114036,
"puct/buffer_value/std": 0.10321993638053985,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5056617183439782,
"puct/buffer_timestep/mean": 10.244444444444444,
"puct/buffer_timestep/std": 6.4983948635389295,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 21.0,
"puct/buffer_construction_len/mean": 1166.638888888889,
"puct/buffer_construction_len/std": 863.3992237654406,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5056664806292557,
"puct/sampled_value/std": 3.7459539612926744e-06,
"puct/sampled_value/min": -1.5056729449928525,
"puct/sampled_value/max": -1.5056617183439782,
"puct/sampled_timestep/mean": 21.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 21.0,
"puct/sampled_timestep/max": 21.0,
"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": 5590.138263225555,
"env/all/ac_tokens_per_turn": 7748.40625,
"env/all/ob_tokens_per_turn": 3260.75,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 3967184,
"env/all/total_ob_tokens": 1669504,
"env/all/time/sampling_mean": 265.37195834796876,
"env/all/time/sampling_max": 362.0066907405853,
"env/all/time/env_step_mean": 2855.639624364674,
"env/all/time/env_step_max": 5224.788479804993,
"env/all/reward/mean": 0.5711764285766573,
"env/all/reward/max": 0.6641603455679572,
"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": 0.5711764285766573,
"env/all/correctness": 0.88671875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.568867677994363,
"env/all/raw_score/min": 1.505660493029354,
"env/all/raw_score/max": 4.5834856518640334,
"env/all/initial_raw_score": -1.5056664806292557,
"env/all/initial_raw_score/min": -1.5056729449928525,
"env/all/initial_raw_score/max": -1.5056617183439782,
"env/all/msg": "Success; raw_score=1.5056609734456798",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nimport cvxpy as cp\nimport copy\nimport random\nfrom scipy.optimize import linprog\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Optimizes a sequence of non-negative coefficients to minimize the evaluation function.\n This version enhances exploration by using a tighter LP constraint selection, hybrid perturbations,\n and more aggressive initial exploration to avoid local minima.\n \"\"\"\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 # Initial diversity with more structured patterns\n n_initial = 100\n initial_sequences = []\n for _ in range(8): # Increase number of initial sequences for better exploration\n if np.random.rand() < 0.25:\n # Cosine-like with normalization\n base = np.cos(np.linspace(0, np.pi, n_initial)) * 0.1 + 0.1\n elif np.random.rand() < 0.25:\n # Uniform with scaling\n base = np.random.uniform(0.01, 0.05, size=n_initial)\n elif np.random.rand() < 0.25:\n # Exponential decay\n base = np.random.exponential(scale=0.01, size=n_initial) * 0.1\n elif np.random.rand() < 0.25:\n # Geometric decay\n base = np.random.geometric(p=0.1, size=n_initial) * 0.1\n elif np.random.rand() < 0.25:\n # Gaussian-like\n base = np.random.normal(loc=0.05, scale=0.02, size=n_initial)\n elif np.random.rand() < 0.25:\n # Alternating high-low pattern\n base = np.zeros(n_initial)\n base[::2] = 0.1 + np.random.normal(0, 0.02, n_initial//2)\n base[1::2] = 0.05 + np.random.normal(0, 0.01, n_initial//2)\n else:\n # Random with uniform scaling\n base = np.random.rand(n_initial)\n sum_base = np.sum(base)\n scale_factor = 0.1 / sum_base if sum_base > 0 else 0.1\n initial_sequences.append([max(0.0, x * scale_factor) for x in base])\n best_sequence = initial_sequences[np.random.choice(len(initial_sequences))]\n \n current_sequence = best_sequence.copy()\n best_score = float('inf')\n random_walk_prob = 0.25 # Increased for better exploration\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using adaptive constraint selection and gradient-guided steps.\"\"\"\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 # Lower threshold to include more constraints for better g_sequence\n threshold = 0.25 * max_b_val\n tight_positions = np.where(conv >= threshold)[0]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n \n # Use scipy's linprog for faster LP solution\n def solve_convolution_lp(f_sequence, rhs, tight_positions):\n n = len(f_sequence)\n if n == 0:\n return None\n \n # Objective: maximize sum(g_j)\n c = [-1.0] * n\n # Constraints: sum_{j} f[i] * g[j] <= rhs for each tight position\n A_ub = []\n b_ub = []\n for k in tight_positions:\n row = np.zeros(n)\n for j in range(n):\n i = k - j\n if 0 <= i < n:\n row[j] = f_sequence[i]\n A_ub.append(row)\n b_ub.append(rhs)\n \n # Bounds: g >= 0\n bounds = [(0.0, 1000.0) for _ in range(n)]\n result = linprog(c, A_ub=A_ub, b_ub=b_ub, bounds=bounds, method='highs')\n if result.success:\n return [max(0.0, x) for x in result.x]\n return None\n \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 # Use line search with gradient information\n def objective(t):\n new_seq = [ (1 - t) * x + t * y for x, y in zip(sequence, g_fun) ]\n return evaluate_sequence(new_seq)\n\n # Find optimal t using golden-section search\n low, high = 0.0, 1.0\n for _ in range(100):\n t1 = low + (high - low) / 3\n t2 = high - (high - low) / 3\n f1 = objective(t1)\n f2 = objective(t2)\n if f1 < f2:\n high = t2\n else:\n low = t1\n best_t = (low + high) / 2\n new_seq = [ (1 - best_t) * x + best_t * y for x, y in zip(sequence, g_fun) ]\n return new_seq\n\n def perturb_sequence(seq):\n \"\"\"Strategic perturbation combining targeted and random changes.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n\n conv = np.convolve(seq, seq)\n top_indices = np.argsort(conv)[-50:]\n indices = np.random.choice(top_indices, size=10, replace=False)\n indices = [min(idx, len(seq)-1) for idx in indices]\n\n # Occasionally perturb random elements\n if np.random.rand() < 0.5:\n random_idx = np.random.randint(0, len(seq))\n indices.append(random_idx)\n \n scale = 0.2 * np.std(seq)\n for idx in indices:\n new_val = max(0.0, seq[idx] + np.random.normal(0, 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 larger-scale, diverse random perturbations.\"\"\"\n scale = 0.3 * 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 # Expand only when sequence length is small\n if len(current_sequence) < 500:\n new_seq = current_sequence + [np.mean(current_sequence)]\n new_seq = [x / np.sum(new_seq) for x in new_seq]\n current_sequence = new_seq\n else:\n current_sequence = perturb_sequence(current_sequence)\n else:\n current_sequence = h_function\n\n # Adjust exploration rate to allow more early search\n if random_walk_prob > 0.1:\n random_walk_prob *= 0.98\n\n # Add controlled random walk for exploration\n if np.random.rand() < random_walk_prob:\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": 265.37195834796876,
"env/all/time/policy/min": 136.2768394947052,
"env/all/time/policy/max": 362.0066907405853,
"env/all/time/env_step": 2855.639624364674,
"env/all/time/env_step/min": 0.009138107299804688,
"env/all/time/env_step/max": 5224.788479804993,
"env/all/time/reward_compute": 5.397014319896698e-07,
"env/all/time/reward_compute/min": 2.644956111907959e-07,
"env/all/time/reward_compute/max": 1.7136335372924805e-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": 0.015820523723959923,
"advantage/min": -1.0,
"advantage/max": 9.246454238891602,
"time/assemble_training_data": 7.818105220794678,
"time/kl_vs_base": 88.63510322570801,
"kl_policy_base": 0.0010070089483633637,
"time/train": 521.4642562866211,
"time/save_checkpoint": 13.11054015159607,
"time/total": 6223.275345802307
}
[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