Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step24
ttt-discover-v3-qwen3-8b-ac1-s2-step24 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 23 (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 23 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 23,
"progress/batch": 23,
"optim/lr": 4e-05,
"progress/done_frac": 0.48,
"puct/buffer_size": 376,
"puct/sampled_size": 8,
"puct/T": 11776,
"puct/scale_last": 0.5741269753189022,
"puct/buffer_value/mean": -1.5337089287189867,
"puct/buffer_value/std": 0.10117278797530675,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.505660493029354,
"puct/buffer_timestep/mean": 10.74468085106383,
"puct/buffer_timestep/std": 6.786933895851247,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 22.0,
"puct/buffer_construction_len/mean": 1159.5478723404256,
"puct/buffer_construction_len/std": 845.498639759934,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5056620961928049,
"puct/sampled_value/std": 9.267235265746569e-07,
"puct/sampled_value/min": -1.5056628567117307,
"puct/sampled_value/max": -1.505660493029354,
"puct/sampled_timestep/mean": 22.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 22.0,
"puct/sampled_timestep/max": 22.0,
"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": 5420.316789388657,
"env/all/ac_tokens_per_turn": 7933.1953125,
"env/all/ob_tokens_per_turn": 3329.25,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4061796,
"env/all/total_ob_tokens": 1704576,
"env/all/time/sampling_mean": 274.3807037086226,
"env/all/time/sampling_max": 393.493275642395,
"env/all/time/env_step_mean": 2648.6870290683582,
"env/all/time/env_step_max": 5073.112332820892,
"env/all/reward/mean": 0.48038002116306755,
"env/all/reward/max": 0.6641605351438402,
"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": 0.48038002116306755,
"env/all/correctness": 0.75,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.6025138200935192,
"env/all/raw_score/min": 1.5056600632583812,
"env/all/raw_score/max": 6.423596339536332,
"env/all/initial_raw_score": -1.5056620961928049,
"env/all/initial_raw_score/min": -1.5056628567117307,
"env/all/initial_raw_score/max": -1.505660493029354,
"env/all/msg": "Success; raw_score=1.5162212876089542",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nimport cvxpy as cp\nimport copy\nimport random\nfrom scipy.optimize import line_search\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 enforcing normalization to reduce complexity,\n introducing more precise convex optimization, dynamic constraint selection, and adaptive perturbation strategies.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Helper function for normalization\n def normalize_sequence(sequence):\n n = len(sequence)\n target_sum = np.sqrt(2 * n)\n current_sum = np.sum(sequence)\n if current_sum == 0:\n return [0.0] * n\n scale = target_sum / current_sum\n new_seq = [max(0.0, min(1000.0, x * scale)) for x in sequence]\n return new_seq\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(4):\n if np.random.rand() < 0.3:\n base = np.cos(np.linspace(0, np.pi, n_initial)) * 0.1 + 0.1\n elif np.random.rand() < 0.3:\n base = np.random.uniform(0.01, 0.05, size=n_initial)\n elif np.random.rand() < 0.3:\n base = np.random.exponential(scale=0.01, size=n_initial) * 0.1\n elif np.random.rand() < 0.3:\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 base = np.random.rand(n_initial)\n sum_base = np.sum(base)\n scale_factor = np.sqrt(2 * n_initial) / sum_base if sum_base > 0 else 1.0\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 current_sequence = normalize_sequence(best_sequence).copy()\n best_score = float('inf')\n random_walk_prob = 0.25 # Moderate exploration rate\n n_initial = len(current_sequence)\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 # Select positions where the convolution is near the maximum\n tight_positions = np.argsort(conv)[::-1][:100]\n \n # Use CVXPY for precise LP solution\n g = cp.Variable(n, nonneg=True)\n objective = cp.Maximize(cp.sum(g))\n constraints = []\n M = max_b_val\n for k in tight_positions:\n row = np.zeros(n)\n for i in range(n):\n j = k - i\n if 0 <= j < n:\n row[i] = sequence[i]\n constraints.append(cp.sum(row * g) <= M)\n \n try:\n prob = cp.Problem(objective, constraints)\n prob.solve(solver=cp.SCS, eps=1e-6)\n if prob.status == cp.OPTIMAL:\n g_fun = [max(0.0, float(g[i])) for i in range(n)]\n sum_g = np.sum(g_fun)\n if sum_g <= 0.0:\n return None\n # Normalize to target sum\n scale = np.sqrt(2 * n) / sum_g\n new_seq = [max(0.0, x * scale) for x in g_fun]\n return new_seq\n except:\n return None\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() < 0.3:\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) < 1000:\n new_seq = current_sequence + [np.mean(current_sequence)]\n new_seq = normalize_sequence(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 # Normalize to target sum\n current_sequence = normalize_sequence(current_sequence)\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 current_sequence = normalize_sequence(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": 274.3807037086226,
"env/all/time/policy/min": 84.25772404670715,
"env/all/time/policy/max": 393.493275642395,
"env/all/time/env_step": 2648.6870290683582,
"env/all/time/env_step/min": 0.007670164108276367,
"env/all/time/env_step/max": 5073.112332820892,
"env/all/time/reward_compute": 3.133900463581085e-07,
"env/all/time/reward_compute/min": 2.0116567611694336e-07,
"env/all/time/reward_compute/max": 5.699694156646729e-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": 0.0170854851603508,
"advantage/min": -1.0,
"advantage/max": 1.8186798095703125,
"time/assemble_training_data": 9.169037580490112,
"time/kl_vs_base": 93.58734822273254,
"kl_policy_base": 0.0009636928443796933,
"time/train": 535.7941946983337,
"time/save_checkpoint": 14.377853631973267,
"time/total": 6075.689382076263
}
[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