Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step13
ttt-discover-v3-qwen3-8b-ac1-step13 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 12 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 10, 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 12 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 12,
"progress/batch": 12,
"optim/lr": 4e-05,
"progress/done_frac": 0.26,
"puct/buffer_size": 200,
"puct/sampled_size": 8,
"puct/T": 6144,
"puct/scale_last": 0.3991191331508912,
"puct/buffer_value/mean": -1.538850426742613,
"puct/buffer_value/std": 0.10827540523784361,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5065039076431792,
"puct/buffer_timestep/mean": 5.24,
"puct/buffer_timestep/std": 3.6141942393844855,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 11.0,
"puct/buffer_construction_len/mean": 1217.655,
"puct/buffer_construction_len/std": 932.568418924317,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5068565017369107,
"puct/sampled_value/std": 0.0003304820928205275,
"puct/sampled_value/min": -1.5071867255611309,
"puct/sampled_value/max": -1.5065039076431792,
"puct/sampled_timestep/mean": 11.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 11.0,
"puct/sampled_timestep/max": 11.0,
"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": 5372.85631608963,
"env/all/ac_tokens_per_turn": 9053.9453125,
"env/all/ob_tokens_per_turn": 3370.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4635620,
"env/all/total_ob_tokens": 1725632,
"env/all/time/sampling_mean": 339.7021804987453,
"env/all/time/sampling_max": 547.899983882904,
"env/all/time/env_step_mean": 2628.8956052064896,
"env/all/time/env_step_max": 4954.023403406143,
"env/all/reward/mean": 0.5504332419964426,
"env/all/reward/max": 0.6638270894975976,
"env/all/reward/min": 0.0,
"env/all/format": 0.998046875,
"env/all/format/min": 0.0,
"env/all/format/max": 1.0,
"env/all/reward": 0.5504332419964426,
"env/all/correctness": 0.89453125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 2.1350475803005637,
"env/all/raw_score/min": 1.5064163683325371,
"env/all/raw_score/max": 111.11111111111111,
"env/all/initial_raw_score": -1.5068565017369107,
"env/all/initial_raw_score/min": -1.5071867255611309,
"env/all/initial_raw_score/max": -1.5065039076431792,
"env/all/msg": "Evaluation timed out after 1100 seconds.",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy import optimize\nimport random\nimport cvxpy as cp\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers that minimizes the evaluation score through a hybrid approach combining enhanced LP solutions, targeted perturbations, and simulated annealing. \n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Initialize using diverse strategies (Gaussian, uniform, structured patterns)\n best_sequence = None\n n = 1000\n if 'height_sequence_1' in globals() and isinstance(height_sequence_1, (list, np.ndarray)):\n best_sequence = list(height_sequence_1)\n else:\n # Use Gaussian-initialized sequence, normalized and clamped\n random_seq = np.random.normal(0.5, 0.2, n)\n random_seq = np.maximum(0.0, np.clip(random_seq, 0, 1000.0))\n random_seq *= np.sqrt(2 * n) / np.sum(random_seq)\n best_sequence = random_seq.tolist()\n \n # Alternative initialization: uniform distribution\n uniform_seq = np.ones(n) * np.sqrt(2 * n) / n\n # Third initialization: structured pattern (peak at center)\n peak_seq = np.zeros(n)\n peak_seq[n//2] = 1000.0\n peak_seq = np.repeat(peak_seq, 10) # Spread out the peak\n peak_seq = np.roll(peak_seq, n//2) # Center the peak\n peak_seq = np.maximum(0.0, peak_seq / np.sum(peak_seq) * np.sqrt(2 * n))\n # Combine all three initializations via averaging\n best_sequence = [x * 0.3 + uniform_seq[i] * 0.3 + peak_seq[i] * 0.4 for i, x in enumerate(best_sequence)]\n curr_sequence = [float(x) for x in best_sequence]\n best_score = evaluate_sequence(curr_sequence)\n\n # Evaluate multiple starting points and keep the best\n initial_candidates = [curr_sequence.copy()]\n for _ in range(3):\n new_seq = np.random.normal(0.5, 0.3, n)\n new_seq = np.maximum(0.0, np.clip(new_seq, 0, 1000.0))\n new_seq *= np.sqrt(2 * n) / np.sum(new_seq)\n initial_candidates.append(new_seq.tolist())\n\n # Evaluate each initial candidate to find the best\n for seq in initial_candidates:\n score = evaluate_sequence(seq)\n if score < best_score:\n best_score = score\n best_sequence = seq.copy()\n curr_sequence = seq.copy()\n \n while time.time() < deadline:\n try:\n # Try to get direction from LP solution\n if np.random.random() < 0.5:\n h_function = get_good_direction_to_move_into(curr_sequence)\n else:\n h_function = get_good_direction_to_move_into(curr_sequence)\n if h_function is not None:\n curr_sequence = h_function\n else:\n # Aggressive targeted perturbations\n idx = perturb_targeted(curr_sequence)\n curr_sequence = perturb_sequence(curr_sequence, idx, 1.5) # Reduced perturbation\n \n except Exception:\n print(\"Evaluation error, skipping update\")\n continue\n\n try:\n curr_score = evaluate_sequence(curr_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = curr_sequence.copy()\n print(f\"New best score: {best_score}\")\n except Exception:\n print(\"Evaluation error, skipping update\")\n\n return [float(max(0.0, x)) for x in best_sequence]\n\ndef get_good_direction_to_move_into(sequence):\n \"\"\"Enhanced direction finding with more robust LP solution and blended strategies using CVXPY.\"\"\"\n n = len(sequence)\n if n == 0:\n return None\n \n sum_a = np.sum(sequence)\n if sum_a <= 0.0:\n return None\n \n # Normalize for LP with dynamic rhs\n normalized_sequence = sequence.copy()\n sum_a_norm = np.sum(normalized_sequence)\n normalized_sequence = [x / sum_a_norm for x in normalized_sequence]\n rhs = np.max(np.convolve(normalized_sequence, normalized_sequence))\n \n # Solve LP to maximize sum(b) s.t. conv(f, b) <= rhs, b >= 0\n try:\n b = cp.Variable(n)\n objective = cp.Maximize(cp.sum(b))\n constraints = []\n for k in range(2 * n - 1):\n conv = 0\n for i in range(n):\n j = k - i\n if 0 <= j < n:\n conv += normalized_sequence[i] * b[j]\n constraints.append(conv <= rhs)\n constraints.append(b >= 0)\n prob = cp.Problem(objective, constraints)\n prob.solve(solver=cp.GLPK, time_limit=60.0)\n if prob.status == cp.OPTIMAL:\n g_fun = b.value\n sum_g = np.sum(g_fun)\n if sum_g > 0.0:\n normalized_g = [x * np.sqrt(2 * n) / sum_g for x in g_fun]\n # Refine blending via grid search with adaptive t\n best_score = float('inf')\n best_t = 0.0\n t_values = np.linspace(0.0, 2.0, 20) # Extended t range for exploration\n for t in t_values:\n new_sequence = [(1 - t) * x + t * y for x, y in zip(sequence, normalized_g)]\n curr_score = evaluate_sequence(new_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_t = t\n new_sequence = [(1 - best_t) * x + best_t * y for x, y in zip(sequence, normalized_g)]\n return new_sequence\n else:\n return None\n else:\n return None\n except:\n pass\n\n # If LP fails, target high-contribution indices\n conv = np.convolve(normalized_sequence, normalized_sequence)\n max_conv_idx = np.argmax(conv)\n contributions = []\n for i in range(n):\n j = max_conv_idx - i\n if 0 <= j < n:\n contributions.append((i, abs(normalized_sequence[i] * normalized_sequence[j])))\n if contributions:\n sorted_contributions = sorted(contributions, key=lambda x: x[1], reverse=True)\n idxs = [sorted_contributions[i][0] for i in range(min(2, len(sorted_contributions)))]\n new_seq = normalized_sequence.copy()\n # Introduce smaller perturbation for more stability\n new_seq[idxs[0]] = max(0.0, new_seq[idxs[0]] - np.random.randn() * 1.0)\n new_seq[idxs[1]] = max(0.0, new_seq[idxs[1]] - np.random.randn() * 1.0)\n return new_seq\n else:\n idx = np.random.randint(0, len(normalized_sequence))\n new_seq = normalized_sequence.copy()\n new_seq[idx] = max(0.0, new_seq[idx] - np.random.randn() * 1.0)\n return new_seq\n\ndef perturb_targeted(sequence, max_conv_idx):\n \"\"\"Targeted perturbation of elements contributing to max_conv_idx with multiple indices.\"\"\"\n n = len(sequence)\n contributions = []\n for i in range(n):\n j = max_conv_idx - i\n if 0 <= j < n:\n contributions.append((i, abs(sequence[i] * sequence[j])))\n if contributions:\n sorted_contributions = sorted(contributions, key=lambda x: x[1], reverse=True)\n # Perturb top 3 contributing indices\n idxs = [sorted_contributions[i][0] for i in range(min(3, len(sorted_contributions)))]\n return idxs\n else:\n return [np.random.randint(0, n) for _ in range(3)]\n\ndef perturb_sequence(sequence, idxs, perturb_amount):\n \"\"\"Perturb multiple indices in sequence with controlled amounts.\"\"\"\n new_seq = sequence.copy()\n for idx in idxs:\n new_seq[idx] = max(0.0, new_seq[idx] - np.random.randn() * perturb_amount)\n new_seq[idx] = min(1000.0, new_seq[idx])\n return new_seq\n```",
"env/all/time/policy": 339.7021804987453,
"env/all/time/policy/min": 165.62680315971375,
"env/all/time/policy/max": 547.899983882904,
"env/all/time/env_step": 2628.8956052064896,
"env/all/time/env_step/min": 0.0071201324462890625,
"env/all/time/env_step/max": 4954.023403406143,
"env/all/time/reward_compute": 5.033798515796661e-07,
"env/all/time/reward_compute/min": 2.1979212760925293e-07,
"env/all/time/reward_compute/max": 2.119690179824829e-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": 0.01406047586351633,
"advantage/min": -1.0,
"advantage/max": 5.687336444854736,
"time/assemble_training_data": 10.732922792434692,
"time/kl_vs_base": 95.42301607131958,
"kl_policy_base": 0.000710771419107914,
"time/train": 603.8498244285583,
"time/save_checkpoint": 34.0274703502655,
"time/total": 6118.994006156921
}
[2026-07-09T06:24:18+00:00] job=1812628 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:30:32+00:00] job=1813127 node=node-22 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:07:08+00:00] job=1813128 node=node-6 ngpu=6 ntrain=2 replicas=4 flash_attn=yes