Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step50
ttt-discover-v3-qwen3-8b-ac1-s2-step50 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 49 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 13, 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 49 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 49,
"progress/batch": 49,
"optim/lr": 4e-05,
"progress/done_frac": 1.0,
"puct/buffer_size": 792,
"puct/sampled_size": 8,
"puct/T": 25088,
"puct/scale_last": 0.5750384121917611,
"puct/buffer_value/mean": -1.518711663402337,
"puct/buffer_value/std": 0.07115386812912164,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5047490561564951,
"puct/buffer_timestep/mean": 23.747474747474747,
"puct/buffer_timestep/std": 14.29087682018729,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 48.0,
"puct/buffer_construction_len/mean": 1075.7449494949494,
"puct/buffer_construction_len/std": 587.9875867065198,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5047496260557844,
"puct/sampled_value/std": 5.319299160267103e-07,
"puct/sampled_value/min": -1.504750157282351,
"puct/sampled_value/max": -1.5047490561564951,
"puct/sampled_timestep/mean": 48.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 48.0,
"puct/sampled_timestep/max": 48.0,
"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": 6277.071914434433,
"env/all/ac_tokens_per_turn": 8492.779296875,
"env/all/ob_tokens_per_turn": 3964.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4348303,
"env/all/total_ob_tokens": 2029888,
"env/all/time/sampling_mean": 326.81840781494975,
"env/all/time/sampling_max": 418.80703997612,
"env/all/time/env_step_mean": 2947.7217473215424,
"env/all/time/env_step_max": 5857.789181470871,
"env/all/reward/mean": 0.5995260311807881,
"env/all/reward/max": 0.6645626681445845,
"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": 0.5995260311807881,
"env/all/correctness": 0.92578125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 6.846298549480907,
"env/all/raw_score/min": 1.5047489744591098,
"env/all/raw_score/max": 2072.0,
"env/all/initial_raw_score": -1.5047496260557847,
"env/all/initial_raw_score/min": -1.504750157282351,
"env/all/initial_raw_score/max": -1.5047490561564951,
"env/all/msg": "Success; raw_score=1.5238371055773723",
"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 Strategy: Focus on peak reduction through adaptive LP constraints, refined perturbation, and directional moves with ECOS solver.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 30 # Reserve 30 seconds for final stabilization\n\n # Initialize from previous best if available\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 initial_sequences = []\n # Generate structured and random initial sequences\n base_len = min(4096, np.random.choice([512, 1024, 2048, 4096]))\n base = np.zeros(base_len)\n if np.random.rand() < 0.6:\n base = np.exp(-np.linspace(0, 1, base_len)**2 * 10) * 0.5\n else:\n base = np.random.rand(base_len) * 0.5 * np.exp(-np.linspace(0, 1, base_len)**2 * 10)\n clamped = [min(1000.0, max(0.0, x)) for x in base]\n sum_clamped = np.sum(clamped)\n target_sum = np.sqrt(2 * base_len)\n scaled = [x * target_sum / sum_clamped for x in clamped] if sum_clamped > 0 else [0.0]*base_len\n initial_sequences.append(scaled)\n best_sequence = scaled\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 perturbation targeting convolution peaks.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n conv = np.convolve(seq, seq)\n top_indices = np.argsort(conv)[::-1]\n indices = np.random.choice(top_indices, size=min(50, len(top_indices)), replace=False)\n indices = [min(idx, len(seq)-1) for idx in indices]\n \n # Use large perturbations on peak positions\n scale = 1.0 * np.std(seq)\n new_seq = seq.copy()\n \n for idx in indices:\n if new_seq[idx] > 0.0:\n # Apply targeted mutation\n if np.random.rand() < 0.5:\n new_seq[idx] *= np.random.uniform(0.5, 0.85)\n else:\n new_seq[idx] += np.random.uniform(-0.1, 0.1)\n \n # Clamp and re-normalize\n clamped = [min(1000.0, max(0.0, x)) for x in new_seq]\n sum_clamped = np.sum(clamped)\n target_sum = np.sqrt(2 * len(clamped))\n if sum_clamped > 0:\n new_seq = [x * target_sum / sum_clamped for x in clamped]\n else:\n new_seq = [0.0 for _ in clamped]\n return new_seq\n\n def random_walk(seq):\n \"\"\"Explores new areas with adaptive perturbation.\"\"\"\n n = len(seq)\n current_sum = np.sum(seq)\n target_sum = np.sqrt(2 * n)\n scale = np.std(seq) * (0.1 + 0.1 * np.random.rand())\n perturbation = np.random.normal(0, scale, size=n)\n new_seq = [seq[i] + perturbation[i] for i in range(n)]\n adjustment = target_sum - np.sum(new_seq)\n new_seq[-1] += adjustment\n clamped = [min(1000.0, max(0.0, x)) for x in new_seq]\n sum_clamped = np.sum(clamped)\n target_sum = np.sqrt(2 * len(clamped))\n if sum_clamped > 0:\n new_seq = [x * target_sum / sum_clamped for x in clamped]\n else:\n new_seq = [0.0 for x in clamped]\n return new_seq\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Optimizes with aggressive threshold lowering and adaptive constraint selection using ECOS.\"\"\"\n n = len(sequence)\n sum_seq = np.sum(sequence)\n if sum_seq <= 0.0:\n return None\n conv = np.convolve(sequence, sequence)\n max_b_val = np.max(conv)\n threshold = max_b_val * 0.98 # Aggressive threshold lowering\n tight_positions = np.where(conv >= threshold)[0]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n \n # Solve LP with tighter threshold for better direction\n def solve_convolution_lp(f_sequence, tight_positions, threshold_scale=0.3):\n n = len(f_sequence)\n if n == 0:\n return None\n g = cp.Variable(n, nonneg=True)\n threshold = cp.Variable(nonneg=True)\n objective = cp.Minimize(threshold)\n constraints = [cp.sum(g) == np.sqrt(2 * n)]\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)\n problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.ECOS, verbose=False, eps=1e-4, lb=0)\n if problem.status == cp.OPTIMAL:\n g_fun = g.value\n if g_fun is not None:\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 g_fun = solve_convolution_lp(sequence, tight_positions)\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 \n # Golden-section search with adaptive iterations\n def objective(t):\n new_seq = [ (1 - t) * x + t * y for x, y in zip(sequence, g_fun) ]\n clamped = [min(1000.0, max(0.0, x)) for x in new_seq]\n sum_clamped = np.sum(clamped)\n target_sum = np.sqrt(2 * len(clamped))\n if sum_clamped > 0:\n new_seq = [x * target_sum / sum_clamped for x in clamped]\n else:\n new_seq = [0.0 for x in clamped]\n return evaluate_sequence(new_seq)\n low, high = 0.0, 1.0\n for _ in range(100): # Increase iteration count for better convergence\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 clamped = [min(1000.0, max(0.0, x)) for x in new_seq]\n sum_clamped = np.sum(clamped)\n target_sum = np.sqrt(2 * len(clamped))\n if sum_clamped > 0:\n new_seq = [x * target_sum / sum_clamped for x in clamped]\n else:\n new_seq = [0.0 for x in clamped]\n return new_seq\n\n def expand_sequence(seq):\n \"\"\"Carefully extends the sequence with new elements.\"\"\"\n n = len(seq)\n if n >= 4096:\n return seq\n new_length = n + 1\n target_sum = np.sqrt(2 * new_length)\n current_sum = np.sum(seq)\n clamped = [min(1000.0, max(0.0, x)) for x in seq]\n sum_clamped = np.sum(clamped)\n if sum_clamped > 0:\n scale = target_sum / sum_clamped\n else:\n scale = 0.5\n new_val = (scale * current_sum) / n * np.random.uniform(1.0, 1.2)\n new_seq = seq + [max(0.0, min(1000.0, new_val))]\n clamped_new = [min(1000.0, max(0.0, x)) for x in new_seq]\n sum_clamped_new = np.sum(clamped_new)\n target_sum_new = np.sqrt(2 * len(clamped_new))\n if sum_clamped_new > 0:\n new_seq = [x * target_sum_new / sum_clamped_new for x in clamped_new]\n else:\n new_seq = [0.0 for x in clamped_new]\n return new_seq\n\n while time.time() < deadline:\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 if len(current_sequence) < 4096:\n current_sequence = expand_sequence(current_sequence)\n else:\n current_sequence = perturb_sequence(current_sequence)\n\n # Controlled random walk for diversity with increased exploration\n if np.random.rand() < 0.01:\n current_sequence = random_walk(current_sequence)\n\n # Targeted adjustment on convolution peaks\n if np.random.rand() < 0.005:\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)[-50:]\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)\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: {curr_score}\")\n except Exception:\n pass\n\n return [float(max(0.0, x)) for x in best_sequence]\n```",
"env/all/time/policy": 326.81840781494975,
"env/all/time/policy/min": 155.91529631614685,
"env/all/time/policy/max": 418.80703997612,
"env/all/time/env_step": 2947.7217473215424,
"env/all/time/env_step/min": 0.98744797706604,
"env/all/time/env_step/max": 5857.789181470871,
"env/all/time/reward_compute": 8.735805749893188e-07,
"env/all/time/reward_compute/min": 2.3096799850463867e-07,
"env/all/time/reward_compute/max": 2.641230821609497e-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": 0.006368670146912336,
"advantage/min": -1.0,
"advantage/max": 0.620266318321228,
"time/assemble_training_data": 7.035703897476196,
"time/kl_vs_base": 100.0637755393982,
"kl_policy_base": 0.0007984894909895957,
"time/train": 602.3830173015594,
"time/save_checkpoint": 25.46283769607544,
"time/total": 7015.448225259781
}
[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