Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step20
ttt-discover-v3-qwen3-8b-ac1-s2-step20 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 19 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 16, 2026
Repo size
354 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 99%
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 19 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 19,
"progress/batch": 19,
"optim/lr": 4e-05,
"progress/done_frac": 0.4,
"puct/buffer_size": 312,
"puct/sampled_size": 8,
"puct/T": 9728,
"puct/scale_last": 0.5740660929084676,
"puct/buffer_value/mean": -1.539455682559602,
"puct/buffer_value/std": 0.11018887644837703,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5057213754397887,
"puct/buffer_timestep/mean": 8.743589743589743,
"puct/buffer_timestep/std": 5.632861224571347,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 18.0,
"puct/buffer_construction_len/mean": 1192.275641025641,
"puct/buffer_construction_len/std": 924.7782310295413,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5057512110704696,
"puct/sampled_value/std": 1.600262359395796e-05,
"puct/sampled_value/min": -1.5057760123327657,
"puct/sampled_value/max": -1.5057213754397887,
"puct/sampled_timestep/mean": 18.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 18.0,
"puct/sampled_timestep/max": 18.0,
"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": 5405.130985736847,
"env/all/ac_tokens_per_turn": 8038.26171875,
"env/all/ob_tokens_per_turn": 3149.75,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4115590,
"env/all/total_ob_tokens": 1612672,
"env/all/time/sampling_mean": 279.8426949023269,
"env/all/time/sampling_max": 362.80132246017456,
"env/all/time/env_step_mean": 2709.5276941508055,
"env/all/time/env_step_max": 5041.998979568481,
"env/all/reward/mean": 0.5324332233033076,
"env/all/reward/max": 0.6641523921081173,
"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": 0.5324332233033076,
"env/all/correctness": 0.85546875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.8602563725440087,
"env/all/raw_score/min": 1.5056785238465065,
"env/all/raw_score/max": 40.65442375254135,
"env/all/initial_raw_score": -1.5057512110704694,
"env/all/initial_raw_score/min": -1.5057760123327657,
"env/all/initial_raw_score/max": -1.5057213754397887,
"env/all/msg": "Success; raw_score=1.5057213842278692",
"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 # Generate diverse initial sequences with more types\n n = 100\n initial_sequences = []\n for _ in range(10):\n if np.random.rand() < 0.1: # Spike\n base = np.zeros(n)\n base[0] = 0.1\n elif np.random.rand() < 0.2: # Cosine-like\n base = np.cos(np.linspace(0, np.pi, n)) * 0.1 + 0.1\n elif np.random.rand() < 0.3: # Uniform\n base = np.random.uniform(0.01, 0.05, size=n)\n elif np.random.rand() < 0.4: # Exponential\n base = np.random.exponential(scale=0.01, size=n)\n elif np.random.rand() < 0.45: # Gaussian-like\n base = np.random.normal(loc=0.05, scale=0.01, size=n)\n base = np.clip(base, 0.005, 0.1)\n elif np.random.rand() < 0.5: # Periodic\n base = np.sin(np.linspace(0, 2 * np.pi, n)) * 0.05 + 0.1\n elif np.random.rand() < 0.55: # Random peak\n base = np.random.rand(n)\n base *= 0.1\n max_idx = np.random.randint(0, n)\n base[max_idx] += 0.05\n else: # Random\n base = np.random.rand(n)\n sum_base = np.sum(base)\n scale_factor = 0.2 / sum_base if sum_base > 0 else 0.2\n initial_sequences.append([max(0.0, x * scale_factor) for x in base])\n # Select the best initial sequence\n best_sequence = initial_sequences[0]\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\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 based on proximity to max_b_val\n threshold = 0.9 * max_b_val # Adjust threshold to 0.9 for more constraints\n tight_positions = np.where(conv >= threshold)[0]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n # Solve LP with enhanced constraint selection\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 # Binary search for optimal t\n best_t = 0.0\n best_score = float('inf')\n t_low, t_high = 0.001, 0.5\n for _ in range(50): # Perform binary search iterations\n t_mid = (t_low + t_high) / 2\n new_seq = [(1 - t_mid) * x + t_mid * y for x, y in zip(sequence, g_fun)]\n try:\n curr_score = evaluate_sequence(new_seq)\n if curr_score < best_score:\n best_score = curr_score\n best_t = t_mid\n except Exception:\n pass\n # Adjust search range based on score\n if best_score < float('inf'):\n t_high = t_mid\n else:\n t_low = t_mid\n if best_score < float('inf'):\n return [(1 - best_t) * x + best_t * y for x, y in zip(sequence, g_fun)]\n else:\n return None\n\n def solve_convolution_lp(f_sequence, rhs, tight_positions):\n \"\"\"Solves LP with enhanced 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 minimizing -sum(g_j)\n objective = cp.Minimize(-cp.sum(g))\n\n # Constraints: ensure max(convolution) <= rhs\n constraints = []\n for k in tight_positions:\n # Coefficients for the constraint: sum_{j=0}^{n-1} f[i] * g[j] <= rhs\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 with a more efficient solver\n problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.SCS, 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 structured strategies for local minima escape.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n\n conv = np.convolve(seq, seq)\n # Select top 30 indices with highest convolution values\n top_indices = np.argsort(conv)[-30:]\n indices = np.random.choice(top_indices, size=min(30, len(top_indices)), replace=False)\n indices = [min(idx, len(seq)-1) for idx in indices]\n # Also, sometimes perturb a random element\n if np.random.rand() < 0.3:\n random_idx = np.random.randint(0, len(seq))\n indices.append(random_idx)\n for idx in indices:\n # Larger random variation with adaptive scaling\n scale = 0.2 * np.std(seq)\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 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 # Expand sequence length if possible\n if len(current_sequence) < 1000:\n n = len(current_sequence)\n new_seq = current_sequence + [0.05] # Add a small value to maintain sum\n new_seq = np.array(new_seq)\n new_seq /= np.sum(new_seq) # Normalize\n current_sequence = new_seq.tolist()\n print(f\"Expanded sequence to length {len(current_sequence)}\")\n else:\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() < 0.2:\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": 279.8426949023269,
"env/all/time/policy/min": 124.96166825294495,
"env/all/time/policy/max": 362.80132246017456,
"env/all/time/env_step": 2709.5276941508055,
"env/all/time/env_step/min": 0.0050122737884521484,
"env/all/time/env_step/max": 5041.998979568481,
"env/all/time/reward_compute": 3.4831464290618896e-07,
"env/all/time/reward_compute/min": 2.2724270820617676e-07,
"env/all/time/reward_compute/max": 9.052455425262451e-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": 0.019451016560196877,
"advantage/min": -1.0,
"advantage/max": 2.997549533843994,
"time/assemble_training_data": 5.497846603393555,
"time/kl_vs_base": 91.44190764427185,
"kl_policy_base": 0.0008949225884862244,
"time/train": 531.266594171524,
"time/save_checkpoint": 17.373159646987915,
"time/total": 6052.799624919891
}
[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