Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step41
ttt-discover-v3-qwen3-8b-ac1-s2-step41 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 40 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 12, 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 40 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 40,
"progress/batch": 40,
"optim/lr": 4e-05,
"progress/done_frac": 0.82,
"puct/buffer_size": 648,
"puct/sampled_size": 8,
"puct/T": 20480,
"puct/scale_last": 0.5750242641398107,
"puct/buffer_value/mean": -1.5218133299952095,
"puct/buffer_value/std": 0.07832652151933005,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5047632042084456,
"puct/buffer_timestep/mean": 19.246913580246915,
"puct/buffer_timestep/std": 11.693124353254769,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 39.0,
"puct/buffer_construction_len/mean": 1092.5771604938273,
"puct/buffer_construction_len/std": 648.8450353117664,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5047766467890407,
"puct/sampled_value/std": 6.9390128343157745e-06,
"puct/sampled_value/min": -1.5047901789627085,
"puct/sampled_value/max": -1.5047632042084456,
"puct/sampled_timestep/mean": 39.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 39.0,
"puct/sampled_timestep/max": 39.0,
"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": 6334.711719989777,
"env/all/ac_tokens_per_turn": 8040.98828125,
"env/all/ob_tokens_per_turn": 4118.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4116986,
"env/all/total_ob_tokens": 2108608,
"env/all/time/sampling_mean": 302.1069725556299,
"env/all/time/sampling_max": 404.64695715904236,
"env/all/time/env_step_mean": 3051.8915588059463,
"env/all/time/env_step_max": 5937.633026599884,
"env/all/reward/mean": 0.5986975745806846,
"env/all/reward/max": 0.6645585853253039,
"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": 0.5986975745806846,
"env/all/correctness": 0.912109375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.8991824423081864,
"env/all/raw_score/min": 1.5047582191191022,
"env/all/raw_score/max": 167.8125256583273,
"env/all/initial_raw_score": -1.5047766467890407,
"env/all/initial_raw_score/min": -1.5047901789627085,
"env/all/initial_raw_score/max": -1.5047632042084456,
"env/all/msg": "Success; raw_score=1.5047629690957858",
"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 \"\"\"\n Optimizes a sequence of non-negative coefficients to minimize the evaluation function.\n This approach focuses on tighter LP constraints, adaptive perturbation scaling, \n and diverse initial patterns to escape local minima and improve convergence.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Try to start from known best construction\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 patterns\n initial_sequences = []\n for _ in range(150): # More diverse initial sequences\n base_len = np.random.choice([200, 400, 600, 800, 1000])\n base = np.zeros(base_len)\n if np.random.rand() < 0.25: # Gauss-like decay\n base = np.exp(-np.linspace(0, 1, base_len)**2 * 10) * 0.1 + 0.02\n elif np.random.rand() < 0.25: # Multi-peak decay\n base = np.zeros_like(base)\n num_peaks = np.random.randint(4, 8)\n for i in range(num_peaks):\n x = np.random.uniform(0, 1)\n base += np.exp(-((np.linspace(0, 1, base_len) - x)**2)*10) * 0.1\n base *= 0.2\n elif np.random.rand() < 0.25: # Random wave with high amplitude\n base = np.sin(np.linspace(0, 4 * np.pi, base_len)) * 0.1 + 0.05\n elif np.random.rand() < 0.25: # Exponential decay\n base = np.random.exponential(scale=0.2, size=base_len) * 0.05\n elif np.random.rand() < 0.1: # Flat sequence\n base = np.ones(base_len) * 0.01\n elif np.random.rand() < 0.1: # Cosine wave\n base = np.cos(np.linspace(0, 2*np.pi, base_len)) * 0.1 + 0.5\n else: # Random but structured\n base = np.random.rand(base_len) * 0.5 + np.exp(-np.linspace(0, 1, base_len)**2 * 10) * 0.1\n sum_base = np.sum(base)\n scale_factor = np.sqrt(2 * base_len) / sum_base if sum_base > 0 else 0.5\n initial_sequences.append([max(0.0, min(1000.0, 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 best_sequence = list(current_sequence)\n\n def perturb_sequence(seq):\n \"\"\"Strategic element-wise adjustments to reduce peak convolution values.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n conv = np.convolve(seq, seq)\n top_indices = np.argsort(conv)[-250:] # Top 250 convolution positions\n indices = np.random.choice(top_indices, size=min(100, len(top_indices)), replace=False)\n indices = [min(idx, len(seq)-1) for idx in indices]\n scale = 0.05 * np.std(seq) # Slightly larger perturbations\n new_seq = seq.copy()\n for idx in indices:\n if new_seq[idx] > 0.0:\n # Compute a small perturbation to this element\n perturbation = np.random.normal(0, scale) * 0.7\n new_val = new_seq[idx] - perturbation\n # Adjust another element to maintain sum\n adjustment = new_val - new_seq[idx]\n other_idx = 0\n if adjustment != 0:\n new_seq[other_idx] += adjustment\n new_seq[idx] = new_val\n # Clamp values to [0, 1000]\n new_seq = [min(1000.0, max(0.0, x)) for x in new_seq]\n return new_seq\n\n def random_walk(seq):\n \"\"\"Small random perturbations to all elements, keeping the sum fixed.\"\"\"\n n = len(seq)\n target_sum = np.sqrt(2 * n)\n current_sum = np.sum(seq)\n scale = 0.05 * np.std(seq)\n # Add small random noise to all elements\n perturbation = np.random.normal(0, scale, size=n)\n new_seq = [seq[i] + perturbation[i] for i in range(n)]\n # Adjust the last element to maintain the sum\n adjustment = target_sum - np.sum(new_seq)\n new_seq[-1] += adjustment\n # Clamp values to [0, 1000]\n new_seq = [min(1000.0, max(0.0, x)) for x in new_seq]\n return new_seq\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Optimizes through dynamic constraint selection and direct blending.\"\"\"\n n = len(sequence)\n sum_sequence = np.sum(sequence)\n if sum_sequence <= 0.0:\n return None\n conv = np.convolve(sequence, sequence)\n max_b_val = np.max(conv)\n threshold = max(0.9 * max_b_val, 0.05 * np.sqrt(n)) if max_b_val > 0 else 0.05\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(sequence, threshold, tight_positions, threshold_scale=0.5)\n if g_fun is None:\n return None\n sum_g = np.sum(g_fun)\n if sum_g <= 0.0:\n return None\n # Golden-section search for optimal t\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 low, high = 0.0, 1.0\n for _ in range(200):\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 solve_convolution_lp(f_sequence, rhs, tight_positions, threshold_scale=0.5):\n n = len(f_sequence)\n if n == 0:\n return None\n g = cp.Variable(n, nonneg=True)\n objective = cp.Minimize(-cp.sum(g))\n constraints = []\n # Use all positions for constraints instead of subset\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) <= threshold_scale * rhs)\n problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.GLPK, verbose=False, eps=1e-6)\n if problem.status == cp.OPTIMAL:\n g_fun = g.value\n # Normalize to have sum equal to sqrt(2n)\n target_sum = np.sqrt(2 * n)\n current_sum = np.sum(g_fun)\n if current_sum > 0:\n g_fun = [x * (target_sum / current_sum) for x in g_fun]\n # Clamp values to [0, 1000]\n g_fun = [min(1000.0, max(0.0, x)) for x in g_fun]\n return g_fun\n else:\n return None\n except:\n return None\n\n def expand_sequence(seq):\n \"\"\"Expand sequence with optimized new element based on current state.\"\"\"\n n = len(seq)\n if n >= 1500:\n return seq\n # Add a new element that balances the convolution\n new_length = n + 1\n target_sum = np.sqrt(2 * new_length)\n current_sum = np.sum(seq)\n new_val = target_sum - current_sum\n new_val = max(0.0, min(1000.0, new_val))\n new_seq = seq + [new_val]\n return new_seq\n\n while time.time() < deadline:\n # LP-based move\n h_function = get_good_direction_to_move_into(current_sequence)\n if h_function is not None:\n current_sequence = h_function\n else:\n # Expand if not already at limit\n if len(current_sequence) < 1500:\n current_sequence = expand_sequence(current_sequence)\n else:\n current_sequence = perturb_sequence(current_sequence)\n\n # Occasionally try random walk for exploration\n if np.random.rand() < 0.1:\n current_sequence = random_walk(current_sequence)\n\n # Try structured balancing on key positions\n if np.random.rand() < 0.02:\n seq = [max(0.0, x) for x in current_sequence]\n n = len(seq)\n conv = np.convolve(seq, seq)\n peak_indices = np.argsort(conv)[-10:] # Focus on important peak positions\n peak_indices = np.array([min(idx, n-1) for idx in peak_indices])\n for idx in peak_indices:\n if seq[idx] > 0.0:\n seq[idx] = max(0.0, seq[idx] * 0.8)\n for shift in [-1, 0, 1]:\n neighbor_idx = idx + shift\n if 0 <= neighbor_idx < n:\n seq[neighbor_idx] = min(1000.0, seq[neighbor_idx] * 1.1)\n current_sequence = [max(0.0, min(1000.0, x)) for x in seq]\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": 302.1069725556299,
"env/all/time/policy/min": 91.47248673439026,
"env/all/time/policy/max": 404.64695715904236,
"env/all/time/env_step": 3051.8915588059463,
"env/all/time/env_step/min": 1.1726129055023193,
"env/all/time/env_step/max": 5937.633026599884,
"env/all/time/reward_compute": 7.208436727523804e-07,
"env/all/time/reward_compute/min": 2.5704503059387207e-07,
"env/all/time/reward_compute/max": 1.9073486328125e-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": 0.016727693378925323,
"advantage/min": -1.0,
"advantage/max": 7.631838798522949,
"time/assemble_training_data": 9.523030519485474,
"time/kl_vs_base": 96.61456084251404,
"kl_policy_base": 0.0008928932365961373,
"time/train": 586.6080045700073,
"time/save_checkpoint": 16.045125484466553,
"time/total": 7048.690118074417
}
[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
[2026-07-12T06:23:00+00:00] job=1827205 node=node-14 ngpu=6 ntrain=2 replicas=4 flash_attn=yes