Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step28
ttt-discover-v3-qwen3-8b-ac1-s2-step28 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 27 (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 27 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 27,
"progress/batch": 27,
"optim/lr": 4e-05,
"progress/done_frac": 0.56,
"puct/buffer_size": 440,
"puct/sampled_size": 8,
"puct/T": 13824,
"puct/scale_last": 0.5741307340340187,
"puct/buffer_value/mean": -1.5296290341855698,
"puct/buffer_value/std": 0.09404712959735324,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5056567343142375,
"puct/buffer_timestep/mean": 12.745454545454546,
"puct/buffer_timestep/std": 7.941188786248755,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 26.0,
"puct/buffer_construction_len/mean": 1136.340909090909,
"puct/buffer_construction_len/std": 783.6142007385864,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5056569086337346,
"puct/sampled_value/std": 1.788171468232634e-07,
"puct/sampled_value/min": -1.505657340648851,
"puct/sampled_value/max": -1.5056567343142375,
"puct/sampled_timestep/mean": 26.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 26.0,
"puct/sampled_timestep/max": 26.0,
"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": 5326.4357788562775,
"env/all/ac_tokens_per_turn": 8347.7734375,
"env/all/ob_tokens_per_turn": 3237.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4274060,
"env/all/total_ob_tokens": 1657344,
"env/all/time/sampling_mean": 297.4386623268947,
"env/all/time/sampling_max": 379.539324760437,
"env/all/time/env_step_mean": 2449.88203490898,
"env/all/time/env_step_max": 4943.948238611221,
"env/all/reward/mean": 0.5035926526380704,
"env/all/reward/max": 0.6641620947772799,
"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": 0.5035926526380704,
"env/all/correctness": 0.771484375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.5366229677090224,
"env/all/raw_score/min": 1.505656527558561,
"env/all/raw_score/max": 2.447070516107621,
"env/all/initial_raw_score": -1.5056569086337344,
"env/all/initial_raw_score/min": -1.505657340648851,
"env/all/initial_raw_score/max": -1.5056567343142375,
"env/all/msg": "Success; raw_score=1.5182479921061118",
"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 approach combines improved constraint selection in LP, adaptive normalization, \n and a hybrid of gradient-based and random search methods to explore the solution space.\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 # Initial sequence with uniform normalization\n initial_n = 1000\n initial_sum = np.sqrt(2 * initial_n)\n initial_sequence = [initial_sum / initial_n] * initial_n\n best_sequence = initial_sequence\n\n current_sequence = best_sequence.copy()\n best_score = float('inf')\n random_walk_prob = 0.25 # High initial exploration probability\n expansion_threshold = 1000 # Maximum expansion limit\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using enhanced LP with precise constraints.\"\"\"\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 tight_positions = np.where(conv == max_b_val)[0]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n\n # Solve LP with precise constraints\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] = sequence[i]\n constraints.append(cp.sum(coeff * g) <= max_b_val)\n\n problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.GLPK, verbose=False, eps=1e-5)\n if problem.status == cp.OPTIMAL:\n return g.value\n except:\n pass\n return None\n\n def perturb_sequence(seq):\n \"\"\"Perturbation with structured strategies to escape local minima.\"\"\"\n n = len(seq)\n if n == 0:\n return None\n\n conv = np.convolve(seq, seq)\n # Select top 50 indices with highest convolution values for targeted changes\n top_indices = np.argsort(conv)[-50:]\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 # 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 return seq\n\n def expand_sequence(seq):\n \"\"\"Strategically expand the sequence to explore new dimensions.\"\"\"\n n = len(seq)\n if n >= expansion_threshold:\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) < expansion_threshold:\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 = [max(0.0, x + np.random.normal(0, 0.1 * np.std(current_sequence))) for x in 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": 297.4386623268947,
"env/all/time/policy/min": 143.95684337615967,
"env/all/time/policy/max": 379.539324760437,
"env/all/time/env_step": 2449.88203490898,
"env/all/time/env_step/min": 0.6506354808807373,
"env/all/time/env_step/max": 4943.948238611221,
"env/all/time/reward_compute": 3.3620744943618774e-07,
"env/all/time/reward_compute/min": 1.8998980522155762e-07,
"env/all/time/reward_compute/max": 1.0579824447631836e-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": 0.013073859736323357,
"advantage/min": -1.0,
"advantage/max": 1.7253222465515137,
"time/assemble_training_data": 5.781349182128906,
"time/kl_vs_base": 84.08424353599548,
"kl_policy_base": 0.0010837916051968932,
"time/train": 553.3765761852264,
"time/save_checkpoint": 18.060568809509277,
"time/total": 5989.930973768234
}
[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