Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step17
ttt-discover-v3-qwen3-8b-ac1-s2-step17 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 16 (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 16 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 16,
"progress/batch": 16,
"optim/lr": 4e-05,
"progress/done_frac": 0.34,
"puct/buffer_size": 264,
"puct/sampled_size": 8,
"puct/T": 8192,
"puct/scale_last": 0.5738888314702828,
"puct/buffer_value/mean": -1.5455562732038788,
"puct/buffer_value/std": 0.11877388122926033,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5058986368779734,
"puct/buffer_timestep/mean": 7.242424242424242,
"puct/buffer_timestep/std": 4.767505561041729,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 15.0,
"puct/buffer_construction_len/mean": 1227.2348484848485,
"puct/buffer_construction_len/std": 1001.381257123837,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.506112228414556,
"puct/sampled_value/std": 9.106530762196331e-05,
"puct/sampled_value/min": -1.5061953510502173,
"puct/sampled_value/max": -1.5058986368779734,
"puct/sampled_timestep/mean": 15.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 15.0,
"puct/sampled_timestep/max": 15.0,
"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": 5245.113119363785,
"env/all/ac_tokens_per_turn": 7859.458984375,
"env/all/ob_tokens_per_turn": 3200.5,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4024043,
"env/all/total_ob_tokens": 1638656,
"env/all/time/sampling_mean": 272.1837897798978,
"env/all/time/sampling_max": 364.7864248752594,
"env/all/time/env_step_mean": 2460.8992701326497,
"env/all/time/env_step_max": 4893.430993318558,
"env/all/reward/mean": 0.5197746964286994,
"env/all/reward/max": 0.664100595379156,
"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": 0.5197746964286994,
"env/all/correctness": 0.8046875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 3.618916685103565,
"env/all/raw_score/min": 1.5057959597040604,
"env/all/raw_score/max": 842.6952151099002,
"env/all/initial_raw_score": -1.5061122284145556,
"env/all/initial_raw_score/min": -1.5061953510502173,
"env/all/initial_raw_score/max": -1.5058986368779734,
"env/all/msg": "Success; raw_score=1.5072738959508",
"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 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 n = 1000\n base = np.ones(n) * (0.1 / n) # Uniform distribution\n best_sequence = [max(0.0, x) for x in base]\n\n current_sequence = best_sequence.copy()\n best_score = float('inf')\n\n # Helper to get better direction using adaptive constraint selection and refined line search\n def get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using adaptive constraint selection and refined line search.\"\"\"\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 # Normalize for the optimization\n normalized_sequence = [x * np.sqrt(2 * n) / sum_sequence for x in sequence]\n conv = np.convolve(normalized_sequence, normalized_sequence)\n max_b_val = np.max(conv)\n\n # Select more positions with tighter tolerance\n tolerance = max_b_val * 0.005 # Slightly relaxed to cover more constraints\n tight_positions = np.where(conv >= max_b_val - tolerance)[0]\n if len(tight_positions) < int(len(conv) * 0.3):\n tight_positions = np.argsort(conv)[::-1][:int(len(conv) * 0.3)]\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n\n # Solve LP with improved solver and constraints\n g_fun = solve_convolution_lp(normalized_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 # Normalize the LP solution\n normalized_g = [x * np.sqrt(2 * n) / sum_g for x in g_fun]\n\n # Line search with more efficient bounds and adaptive step\n low, high = 0.1, 0.8\n best_t = 0.01\n best_score = float('inf')\n for _ in range(300): # More iterations with dynamic bounds\n mid = (low + high) / 2\n new_seq = [(1 - mid)*x + mid*y for x, y in zip(sequence, normalized_g)]\n # Evaluate with safety\n try:\n curr_score = evaluate_sequence(new_seq)\n if curr_score < best_score:\n best_score = curr_score\n best_t = mid\n # Dynamic bounds based on improvement\n if curr_score < best_score:\n low = mid\n else:\n high = mid\n except Exception:\n pass\n\n return [(1 - best_t)*x + best_t*y for x, y in zip(sequence, normalized_g)]\n\n # Solve LP with a more efficient solver\n def solve_convolution_lp(f_sequence, rhs, tight_positions):\n \"\"\"Solves LP with dynamic constraint selection using efficient solver.\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n g = cp.Variable(n, nonneg=True)\n objective = cp.Minimize(-cp.sum(g)) # Maximize sum(g)\n\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) <= rhs)\n\n problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.ECOS, verbose=False)\n if problem.status == cp.OPTIMAL:\n return g.value\n except:\n return None\n return None\n\n # Enhanced perturbation that targets high-contribution elements\n def perturb_sequence(seq):\n conv = np.convolve(seq, seq)\n top_indices = np.argsort(conv)[-50:] # Focus on top 50 contributing indices\n\n # Prioritize elements involved in top convolutions\n indices = np.random.choice(top_indices, size=min(5, len(top_indices)), replace=False)\n indices = [min(idx, len(seq)-1) for idx in indices]\n\n for idx in indices:\n scale = 0.4 * np.std(seq)\n delta = np.random.uniform(-scale, scale)\n new_val = max(0.0, seq[idx] + delta)\n seq[idx] = new_val\n return seq\n\n # Initial aggressive perturbation to shake out local minima\n current_sequence = perturb_sequence(current_sequence)\n\n while time.time() < deadline:\n if np.random.rand() > 0.2: # 80% chance to use LP improvement\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 # 20% chance to use perturbation\n current_sequence = perturb_sequence(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": 272.1837897798978,
"env/all/time/policy/min": 111.87502956390381,
"env/all/time/policy/max": 364.7864248752594,
"env/all/time/env_step": 2460.8992701326497,
"env/all/time/env_step/min": 0.00761103630065918,
"env/all/time/env_step/max": 4893.430993318558,
"env/all/time/reward_compute": 4.731118679046631e-07,
"env/all/time/reward_compute/min": 1.7881393432617188e-07,
"env/all/time/reward_compute/max": 1.4007091522216797e-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": 0.019307700917124748,
"advantage/min": -1.0,
"advantage/max": 6.706753730773926,
"time/assemble_training_data": 5.6921021938323975,
"time/kl_vs_base": 84.98979210853577,
"kl_policy_base": 0.0008503853459842503,
"time/train": 523.9534976482391,
"time/save_checkpoint": 16.634709358215332,
"time/total": 5884.352871417999
}
[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