Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step40
ttt-discover-v3-qwen3-8b-ac1-s2-step40 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 39 (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 39 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 39,
"progress/batch": 39,
"optim/lr": 4e-05,
"progress/done_frac": 0.8,
"puct/buffer_size": 632,
"puct/sampled_size": 8,
"puct/T": 19968,
"puct/scale_last": 0.5749972788666631,
"puct/buffer_value/mean": -1.5222442592028533,
"puct/buffer_value/std": 0.0792643721141547,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5047901894815932,
"puct/buffer_timestep/mean": 18.746835443037973,
"puct/buffer_timestep/std": 11.404494288901162,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 38.0,
"puct/buffer_construction_len/mean": 1094.9208860759493,
"puct/buffer_construction_len/std": 656.8376025670656,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5048010303031476,
"puct/sampled_value/std": 9.012863984938512e-06,
"puct/sampled_value/min": -1.504809956279746,
"puct/sampled_value/max": -1.5047901894815932,
"puct/sampled_timestep/mean": 38.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 38.0,
"puct/sampled_timestep/max": 38.0,
"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": 6250.329231262207,
"env/all/ac_tokens_per_turn": 8132.271484375,
"env/all/ob_tokens_per_turn": 4114.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4163723,
"env/all/total_ob_tokens": 2106368,
"env/all/time/sampling_mean": 303.31921765068546,
"env/all/time/sampling_max": 397.6939551830292,
"env/all/time/env_step_mean": 3079.381150621455,
"env/all/time/env_step_max": 5862.214788198471,
"env/all/reward/mean": 0.6241347560964506,
"env/all/reward/max": 0.6645563837271451,
"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": 0.6241347560964506,
"env/all/correctness": 0.951171875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.8027752201570066,
"env/all/raw_score/min": 1.5047632042084456,
"env/all/raw_score/max": 68.5796780555553,
"env/all/initial_raw_score": -1.5048010303031476,
"env/all/initial_raw_score/min": -1.504809956279746,
"env/all/initial_raw_score/max": -1.5047901894815932,
"env/all/msg": "Success; raw_score=1.5047901855002244",
"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 refines the previous method by tightening LP constraints and increasing exploration.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Initialize from the best known 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 initial sequences with varied structures\n initial_sequences = []\n for _ in range(30): # Reduced to focus on quality over quantity\n base_len = np.random.choice([256, 512, 1024])\n base = np.zeros(base_len)\n if np.random.rand() < 0.2: # Gauss-like decay\n base = np.exp(-np.linspace(0, 1, base_len)**2 * 10) * 0.1 + 0.02\n elif np.random.rand() < 0.2: # Multi-peak decay\n base = np.zeros_like(base)\n num_peaks = np.random.randint(3, 7)\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() < 0.2: # Exponential decay\n base = np.random.exponential(scale=0.2, size=base_len) * 0.05\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 target_sum = np.sqrt(2 * base_len)\n if sum_base > 0:\n scale_factor = target_sum / sum_base\n else:\n scale_factor = target_sum / 0.5 # default if sum_base is zero\n scaled_base = [max(0.0, min(1000.0, x * scale_factor)) for x in base]\n initial_sequences.append(scaled_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 \"\"\"Focused perturbations on top convolution positions.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n conv = np.convolve(seq, seq)\n top_indices = np.argsort(conv)[-1000:] # Increased top positions\n indices = np.random.choice(top_indices, size=min(60, len(top_indices)), replace=False)\n indices = [min(idx, len(seq)-1) for idx in indices]\n scale = 0.05 * np.std(seq) # Reduced perturbation scale\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 \"\"\"Small, adaptive perturbations to all elements.\"\"\"\n scale = 0.08 * 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 def get_good_direction_to_move_into(sequence):\n \"\"\"Optimizes through dynamic constraint selection and direct blending with refined thresholds.\"\"\"\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_scale = 0.2 # Tightened threshold\n threshold = max(0.25 * max_b_val, 0.02 * max_b_val * np.sqrt(n)) if max_b_val > 0 else 0.02\n # Select top 200 positions with highest convolution values\n tight_positions = np.argsort(conv)[-200:]\n tight_positions = np.unique(tight_positions)\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=threshold_scale)\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(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 # Normalize\n new_seq = [max(0.0, x) for x in new_seq]\n target_sum = np.sqrt(2 * len(new_seq))\n if np.sum(new_seq) > 0:\n scale_factor = target_sum / np.sum(new_seq)\n else:\n scale_factor = target_sum / 0.5 # default if sum is zero\n new_seq = [x * scale_factor for x in new_seq]\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 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 return g.value if problem.status == cp.OPTIMAL else None\n except:\n return None\n\n def expand_sequence(seq):\n \"\"\"Expand sequence with line search to find optimal new element.\"\"\"\n n = len(seq)\n if n >= 1500:\n return seq\n current_sum = np.sum(seq)\n conv = np.convolve(seq, seq)\n max_b = np.max(conv)\n def eval_func(x):\n new_seq = seq + [x]\n new_sum = current_sum + x\n # Normalize\n target_sum = np.sqrt(2 * (n + 1))\n if new_sum > 0:\n scale_factor = target_sum / new_sum\n else:\n scale_factor = target_sum / 0.5\n new_seq_scaled = [y * scale_factor for y in new_seq]\n new_conv = np.convolve(new_seq_scaled, new_seq_scaled)\n new_max_b = np.max(new_conv)\n if new_sum < 0.01:\n return float('inf')\n return 2 * (n + 1) * new_max_b / (new_sum ** 2)\n # Line search around the current last element's value\n initial_x = seq[-1] if n > 0 else 0.01\n best_x = initial_x\n best_score = eval_func(best_x)\n # Try a range of x values around initial_x\n for delta in [ -0.2, -0.15, -0.1, -0.05, 0.0, 0.05, 0.1, 0.15, 0.2 ]:\n x_candidate = initial_x + delta\n x_candidate = max(0.0, x_candidate)\n score = eval_func(x_candidate)\n if score < best_score:\n best_score = score\n best_x = x_candidate\n # Try additional fixed values\n for x_candidate in [0.01, 0.02, 0.05, 0.1, 0.15, 0.2]:\n score = eval_func(x_candidate)\n if score < best_score:\n best_score = score\n best_x = x_candidate\n new_seq = seq + [best_x]\n # Normalize\n new_sum = np.sum(new_seq)\n target_sum = np.sqrt(2 * (n + 1))\n if new_sum > 0:\n scale_factor = target_sum / new_sum\n else:\n scale_factor = target_sum / 0.5\n new_seq = [x * scale_factor for x in new_seq]\n return new_seq\n\n def structured_balancing(seq):\n \"\"\"Targeted adjustments on peak positions and their neighbors.\"\"\"\n seq = [max(0.0, x) for x in seq]\n n = len(seq)\n conv = np.convolve(seq, seq)\n peak_indices = np.argsort(conv)[-15:] # Focus on critical 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] * 0.95) # Reduce peak more gently\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) # Increase neighbors\n return [max(0.0, min(1000.0, x)) for x in 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 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() < 0.05:\n current_sequence = random_walk(current_sequence)\n\n # Apply structured balancing to refine\n if np.random.rand() < 0.03:\n current_sequence = structured_balancing(current_sequence)\n\n try:\n # Normalize current_sequence\n target_sum = np.sqrt(2 * len(current_sequence))\n if np.sum(current_sequence) > 0:\n scale_factor = target_sum / np.sum(current_sequence)\n else:\n scale_factor = target_sum / 0.5\n current_sequence = [max(0.0, x * scale_factor) for x in current_sequence]\n # Clip to 0-1000\n current_sequence = [min(1000.0, max(0.0, x)) for x in current_sequence]\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": 303.31921765068546,
"env/all/time/policy/min": 131.41855716705322,
"env/all/time/policy/max": 397.6939551830292,
"env/all/time/env_step": 3079.381150621455,
"env/all/time/env_step/min": 0.8766982555389404,
"env/all/time/env_step/max": 5862.214788198471,
"env/all/time/reward_compute": 5.206093192100525e-07,
"env/all/time/reward_compute/min": 1.9744038581848145e-07,
"env/all/time/reward_compute/max": 2.4884939193725586e-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": 0.05276833474636078,
"advantage/min": -1.0,
"advantage/max": 21.222938537597656,
"time/assemble_training_data": 5.729107856750488,
"time/kl_vs_base": 88.97395372390747,
"kl_policy_base": 0.0009551941766403615,
"time/train": 591.0017492771149,
"time/save_checkpoint": 19.655835390090942,
"time/total": 6957.982592344284
}
[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