Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step27
ttt-discover-v3-qwen3-8b-ac1-s2-step27 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 26 (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 26 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 26,
"progress/batch": 26,
"optim/lr": 4e-05,
"progress/done_frac": 0.54,
"puct/buffer_size": 424,
"puct/sampled_size": 8,
"puct/T": 13312,
"puct/scale_last": 0.5741301843708868,
"puct/buffer_value/mean": -1.530533613689694,
"puct/buffer_value/std": 0.09568766233295263,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5056572839773694,
"puct/buffer_timestep/mean": 12.245283018867925,
"puct/buffer_timestep/std": 7.652612478761028,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 25.0,
"puct/buffer_construction_len/mean": 1141.4858490566037,
"puct/buffer_construction_len/std": 797.8063821408401,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5056581786440155,
"puct/sampled_value/std": 7.45129549017504e-07,
"puct/sampled_value/min": -1.5056592668764464,
"puct/sampled_value/max": -1.5056572839773694,
"puct/sampled_timestep/mean": 25.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 25.0,
"puct/sampled_timestep/max": 25.0,
"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": 5341.9471690654755,
"env/all/ac_tokens_per_turn": 8390.302734375,
"env/all/ob_tokens_per_turn": 3317.125,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4295835,
"env/all/total_ob_tokens": 1698368,
"env/all/time/sampling_mean": 305.89713480416685,
"env/all/time/sampling_max": 411.85002613067627,
"env/all/time/env_step_mean": 2556.159215304069,
"env/all/time/env_step_max": 4968.886510848999,
"env/all/reward/mean": 0.5046396681674902,
"env/all/reward/max": 0.6641620035750296,
"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": 0.5046396681674902,
"env/all/correctness": 0.775390625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.5743460661538582,
"env/all/raw_score/min": 1.5056567343142375,
"env/all/raw_score/max": 13.947540106149127,
"env/all/initial_raw_score": -1.5056581786440155,
"env/all/initial_raw_score/min": -1.5056592668764464,
"env/all/initial_raw_score/max": -1.5056572839773694,
"env/all/msg": "Success; raw_score=1.5056567343278038",
"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 normalizes the sequence to have a fixed sum (sqrt(2n)) and focuses on minimizing the max convolution.\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 a more structured initial sequence with a decaying pattern and normalization\n initial_n = 500\n base = np.random.exponential(scale=0.05, size=initial_n)\n sum_base = np.sum(base)\n if sum_base > 0:\n scale_factor = np.sqrt(2 * initial_n) / sum_base\n else:\n scale_factor = 1.0\n initial_sequence = [max(0.0, x * scale_factor) for x in base]\n best_sequence = initial_sequence.copy()\n \n current_sequence = best_sequence.copy()\n best_score = float('inf')\n random_walk_prob = 0.15 # Higher initial probability for exploration\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using adaptive constraint selection with dynamic threshold.\"\"\"\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 # Aggressive constraint selection with dynamic threshold\n threshold = max(0.05 * max_b_val, 0.01 * max_b_val * (1 + np.log(n + 1)))\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 and solver adjustments\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 # Use golden-section search on t to find optimal step\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\n # Find optimal t between 0 and 1\n low = 0.0\n high = 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 return new_seq\n\n def solve_convolution_lp(f_sequence, rhs, tight_positions):\n \"\"\"Solves LP with enhanced constraint selection and solver adjustments.\"\"\"\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 # Include all tight positions, more constraints for stronger optimization\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.GLPK, verbose=False, eps=1e-6)\n if problem.status == cp.OPTIMAL:\n g_fun = g.value\n sum_g = np.sum(g_fun)\n if sum_g > 0:\n scale_factor = np.sqrt(2 * n) / sum_g\n g_fun_normalized = [x * scale_factor for x in g_fun]\n else:\n g_fun_normalized = g_fun\n return g_fun_normalized\n except:\n pass\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 75 indices with highest convolution values for targeted changes\n top_indices = np.argsort(conv)[-75:]\n indices = np.random.choice(top_indices, size=min(75, len(top_indices)), replace=False)\n indices = [min(idx, len(seq)-1) for idx in indices]\n # Occasionally perturb a random element\n if np.random.rand() < 0.2:\n random_idx = np.random.randint(0, len(seq))\n indices.append(random_idx)\n # Use adaptive scaling based on sequence properties\n scale = 0.1 * np.std(seq)\n for idx in indices:\n # Add random variation with varying scale\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\n # Ensure sum remains sqrt(2n)\n sum_seq = np.sum(seq)\n target_sum = np.sqrt(2 * n)\n if sum_seq != 0:\n scale_factor = target_sum / sum_seq\n seq = [x * scale_factor for x in seq]\n\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 # Ensure sum remains sqrt(2n)\n sum_seq = np.sum(new_seq)\n target_sum = np.sqrt(2 * len(seq))\n if sum_seq != 0:\n scale_factor = target_sum / sum_seq\n new_seq = [x * scale_factor for x in new_seq]\n return new_seq\n\n def expand_sequence(seq):\n \"\"\"Strategically expand the sequence to explore new dimensions.\"\"\"\n n = len(seq)\n if n >= 1500: # Increase limit for better exploration\n return seq.copy()\n # Add a new element with value similar to the spread of current elements\n val = np.mean(seq) + np.random.normal(0, 0.05) * np.std(seq)\n val = max(0.0, min(1000.0, val))\n new_seq = seq + [val]\n # Normalize to ensure sum remains >= 0.01\n new_sum = np.sum(new_seq)\n if new_sum > 0.0:\n new_seq = [x / new_sum for x in new_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 if possible\n if len(current_sequence) < 1500:\n current_sequence = expand_sequence(current_sequence)\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 # Adjust random walk probability to explore less as solution improves\n if random_walk_prob > 0.05:\n random_walk_prob *= 0.995\n\n # Occasionally try random walk for exploration\n if np.random.rand() < random_walk_prob:\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": 305.89713480416685,
"env/all/time/policy/min": 124.35803270339966,
"env/all/time/policy/max": 411.85002613067627,
"env/all/time/env_step": 2556.159215304069,
"env/all/time/env_step/min": 0.7845690250396729,
"env/all/time/env_step/max": 4968.886510848999,
"env/all/time/reward_compute": 4.0652230381965637e-07,
"env/all/time/reward_compute/min": 1.6391277313232422e-07,
"env/all/time/reward_compute/max": 1.1995434761047363e-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": 0.014850057661533356,
"advantage/min": -1.0,
"advantage/max": 1.6435351371765137,
"time/assemble_training_data": 8.7568678855896,
"time/kl_vs_base": 96.27124691009521,
"kl_policy_base": 0.0009807940805330873,
"time/train": 561.4992139339447,
"time/save_checkpoint": 25.29717707633972,
"time/total": 6035.501905918121
}
[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