Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step33
ttt-discover-v3-qwen3-8b-ac1-s2-step33 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 32 (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 32 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 32,
"progress/batch": 32,
"optim/lr": 4e-05,
"progress/done_frac": 0.66,
"puct/buffer_size": 520,
"puct/sampled_size": 8,
"puct/T": 16384,
"puct/scale_last": 0.5746106261135429,
"puct/buffer_value/mean": -1.525939975649447,
"puct/buffer_value/std": 0.08694231002590075,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5051768422347134,
"puct/buffer_timestep/mean": 15.246153846153845,
"puct/buffer_timestep/std": 9.384161401377355,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 31.0,
"puct/buffer_construction_len/mean": 1115.3653846153845,
"puct/buffer_construction_len/std": 722.4968147080743,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5055950766713564,
"puct/sampled_value/std": 0.00015807775849673706,
"puct/sampled_value/min": -1.5056548292437302,
"puct/sampled_value/max": -1.5051768422347134,
"puct/sampled_timestep/mean": 31.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 31.0,
"puct/sampled_timestep/max": 31.0,
"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": 5532.5161921978,
"env/all/ac_tokens_per_turn": 8302.607421875,
"env/all/ob_tokens_per_turn": 3687.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4250935,
"env/all/total_ob_tokens": 1887936,
"env/all/time/sampling_mean": 548.0373502522707,
"env/all/time/sampling_max": 748.9838359355927,
"env/all/time/env_step_mean": 2690.0380994714797,
"env/all/time/env_step_max": 4776.883606910706,
"env/all/reward/mean": 0.5514598753317499,
"env/all/reward/max": 0.6644962124447896,
"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": 0.5514598753317499,
"env/all/correctness": 0.841796875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.5321190438881176,
"env/all/raw_score/min": 1.5048994631224086,
"env/all/raw_score/max": 3.0265220192847746,
"env/all/initial_raw_score": -1.5055950766713568,
"env/all/initial_raw_score/min": -1.5056548292437302,
"env/all/initial_raw_score/max": -1.5051768422347134,
"env/all/msg": "Success; raw_score=1.5051523831634117",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nimport cvxpy as cp\nimport copy\nimport random\nimport math\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 algorithm uses a hybrid approach combining genetic diversity, adaptive perturbation,\n and structured LP-based optimization to escape local minima and find superior results.\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 # Generate diverse initial sequences with high-sum structures and varied patterns\n initial_sequences = []\n base_len = 128 # Adjusted to a balance between exploration and speed\n for _ in range(150): # More initial sequences for broader coverage\n base = np.zeros(base_len)\n if np.random.rand() < 0.05: # Gauss-like decay\n base = np.exp(-np.linspace(0, 1, base_len)**2) * 0.1\n elif np.random.rand() < 0.05: # Uniform\n base = np.random.uniform(0.05, 0.3, size=base_len)\n elif np.random.rand() < 0.05: # Exponential decay\n base = np.random.exponential(scale=0.2, size=base_len) * 0.05\n elif np.random.rand() < 0.05: # Periodic\n base = np.sin(np.linspace(0, 4 * np.pi, base_len)) * 0.05 + 0.05\n elif np.random.rand() < 0.05: # Random with higher values\n base = np.random.rand(base_len) * 0.5\n elif np.random.rand() < 0.05: # Multi-peak decay\n base = np.zeros_like(base)\n for i in range(5):\n if i % 2 == 0:\n base += np.exp(-((np.linspace(0, 1, base_len) - i*0.3)**2)*10) * 0.1\n else:\n base += np.exp(-((np.linspace(0, 1, base_len) - i*0.3)**2)*30) * 0.03\n base *= 0.05\n else: # Smooth decay with noise\n base = np.random.rand(base_len) * 0.5 + \\\n np.exp(-np.linspace(0, 1, base_len)**2 * 10) * 0.1\n sum_base = np.sum(base)\n scale_factor = 0.5 / (sum_base if sum_base > 0.0 else 0.5)\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:\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 # Enhanced perturbation functions with adaptive strategies\n def adaptive_perturb(sequence):\n \"\"\"Adaptive perturbation combining local and global exploration.\"\"\"\n n = len(sequence)\n if n == 0:\n return None\n conv = np.convolve(sequence, sequence)\n top_indices = np.argsort(conv)[-200:] # Top 200 indices for convolution peaks\n indices = np.random.choice(top_indices, size=min(100, len(top_indices)), replace=False)\n indices = [min(idx, len(sequence)-1) for idx in indices]\n if np.random.rand() < 0.1: # Occasionally perturb a random element\n random_idx = np.random.randint(0, len(sequence))\n indices.append(random_idx)\n scale = 0.3 * np.std(sequence) # Larger perturbation\n for idx in indices:\n new_val = max(0.0, sequence[idx] + np.random.normal(0, scale))\n sequence = [new_val if i == idx else sequence[i] for i in range(len(sequence))]\n return sequence\n\n def adaptive_walk(sequence):\n \"\"\"Adaptive random walk for better exploration balance.\"\"\"\n scale = 0.15 * np.std(sequence) # Increased perturbation scale\n new_seq = [max(0.0, x + np.random.normal(0, scale)) for x in sequence]\n return new_seq\n\n def solve_convolution_lp(f_sequence, rhs, tight_positions):\n \"\"\"Optimized LP solver for directional movement.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n g = cp.Variable(n, nonneg=True)\n objective = cp.Maximize(cp.sum(g))\n constraints = []\n for k in tight_positions:\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 problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.SCS, verbose=False, eps=1e-6)\n return g.value if problem.status == cp.OPTIMAL else None\n except:\n return None\n\n def structured_adjust(sequence):\n \"\"\"Balances values around convolution peaks for better control.\"\"\"\n n = len(sequence)\n conv = np.convolve(sequence, sequence)\n peak_indices = np.argsort(conv)[-10:]\n indices = [min(idx, n-1) for idx in peak_indices]\n for idx in indices:\n if sequence[idx] > 0.0:\n sequence[idx] = max(0.0, sequence[idx] * 0.7)\n for shift in [-1, 0, 1]:\n neighbor_idx = idx + shift\n if 0 <= neighbor_idx < n:\n sequence[neighbor_idx] = min(1000.0, sequence[neighbor_idx] * 1.2)\n sequence = [max(0.0, min(1000.0, x)) for x in sequence]\n return sequence\n\n while time.time() < deadline:\n # Try LP-based move towards better direction\n h_function = None\n try:\n n = len(current_sequence)\n sum_sequence = np.sum(current_sequence)\n if sum_sequence <= 0.0:\n h_function = None\n else:\n conv = np.convolve(current_sequence, current_sequence)\n max_b_val = np.max(conv)\n threshold = max(0.5 * max_b_val, 0.1 * max_b_val * np.sqrt(n)) if max_b_val > 0 else 0.1\n tight_positions = np.where(conv >= threshold)[0]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n g_fun = solve_convolution_lp(current_sequence, max_b_val, tight_positions)\n if g_fun is not None and np.sum(g_fun) > 0.0:\n h_function = [max(0.0, x) for x in g_fun]\n h_function = [min(1000.0, x) for x in h_function]\n except:\n h_function = None\n\n if h_function is not None:\n # Blend current and LP-generated sequence\n current_sequence = [ (1 - 0.2) * x + 0.2 * y for x, y in zip(current_sequence, h_function) ]\n else:\n # Expand sequence if possible\n if len(current_sequence) < 1500:\n n = len(current_sequence)\n new_val = max(0.0, np.max(current_sequence) * 0.8)\n new_seq = current_sequence + [new_val]\n new_sum = np.sum(new_seq)\n if new_sum < 0.01:\n scale_factor = 0.01 / new_sum\n new_seq = [x * scale_factor for x in new_seq]\n current_sequence = new_seq\n else:\n current_sequence = adaptive_perturb(current_sequence)\n\n # Random walk for exploration\n if np.random.rand() < 0.05:\n current_sequence = adaptive_walk(current_sequence)\n\n # Structured balance around peaks\n if np.random.rand() < 0.03:\n current_sequence = structured_adjust(current_sequence)\n\n # Try to balance values across the entire sequence\n if np.random.rand() < 0.02:\n n = len(current_sequence)\n conv = np.convolve(current_sequence, current_sequence)\n avg_conv = np.mean(conv)\n for i in range(n):\n if current_sequence[i] > 0.0:\n current_sequence[i] = max(0.0, current_sequence[i] * 0.9 + avg_conv * 0.1 / 1000)\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\"[INFO] New best score: {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": 548.0373502522707,
"env/all/time/policy/min": 219.8330101966858,
"env/all/time/policy/max": 748.9838359355927,
"env/all/time/env_step": 2690.0380994714797,
"env/all/time/env_step/min": 0.5662517547607422,
"env/all/time/env_step/max": 4776.883606910706,
"env/all/time/reward_compute": 4.745088517665863e-07,
"env/all/time/reward_compute/min": 2.123415470123291e-07,
"env/all/time/reward_compute/max": 1.2405216693878174e-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": 0.009000423364341259,
"advantage/min": -1.0,
"advantage/max": 3.9403204917907715,
"time/assemble_training_data": 5.521830081939697,
"time/kl_vs_base": 146.49741959571838,
"kl_policy_base": 0.0010361558524891734,
"time/train": 1158.403118133545,
"time/save_checkpoint": 20.039142370224,
"time/total": 6864.682686328888
}
[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
[2026-07-11T14:04:05+00:00] job=1824338 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=yes