Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step18
ttt-discover-v3-qwen3-8b-ac1-s2-step18 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 17 (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 17 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 17,
"progress/batch": 17,
"optim/lr": 4e-05,
"progress/done_frac": 0.36,
"puct/buffer_size": 280,
"puct/sampled_size": 8,
"puct/T": 8704,
"puct/scale_last": 0.5739915086441958,
"puct/buffer_value/mean": -1.5432979384361367,
"puct/buffer_value/std": 0.11569468221011266,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5057959597040604,
"puct/buffer_timestep/mean": 7.742857142857143,
"puct/buffer_timestep/std": 5.055932058443017,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 16.0,
"puct/buffer_construction_len/mean": 1214.25,
"puct/buffer_construction_len/std": 973.7790312928875,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5059054280702242,
"puct/sampled_value/std": 5.626569720268235e-05,
"puct/sampled_value/min": -1.5059815844826825,
"puct/sampled_value/max": -1.5057959597040604,
"puct/sampled_timestep/mean": 16.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 16.0,
"puct/sampled_timestep/max": 16.0,
"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": 5314.2210030555725,
"env/all/ac_tokens_per_turn": 7808.84375,
"env/all/ob_tokens_per_turn": 3135.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 3998128,
"env/all/total_ob_tokens": 1605312,
"env/all/time/sampling_mean": 265.6219766661525,
"env/all/time/sampling_max": 348.4655876159668,
"env/all/time/env_step_mean": 2256.6698965686373,
"env/all/time/env_step_max": 4969.078485250473,
"env/all/reward/mean": 0.4752588502899199,
"env/all/reward/max": 0.6641201331041209,
"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": 0.4752588502899199,
"env/all/correctness": 0.740234375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.6381002794305999,
"env/all/raw_score/min": 1.505751660749635,
"env/all/raw_score/max": 17.91113491188795,
"env/all/initial_raw_score": -1.5059054280702244,
"env/all/initial_raw_score/min": -1.5059815844826825,
"env/all/initial_raw_score/max": -1.5057959597040604,
"env/all/msg": "Success; raw_score=1.6010283427761796",
"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 = 500 # Initial sequence size\n initial_sequences = []\n for _ in range(5): # More initial sequences for diversity\n if np.random.rand() < 0.4:\n base = np.cos(np.linspace(0, np.pi, n)) * 0.2 + 0.1\n elif np.random.rand() < 0.6:\n base = np.random.exponential(scale=0.02, size=n)\n else:\n base = np.power(np.linspace(0.1, 0.9, n), 2)\n seq = [max(0.0, x) for x in base]\n initial_sequences.append(seq)\n best_sequence = initial_sequences[0].copy()\n for seq in initial_sequences[1:]:\n curr_score = evaluate_sequence(seq)\n best_score = evaluate_sequence(best_sequence)\n if curr_score < best_score:\n best_sequence = seq.copy()\n \n current_sequence = best_sequence.copy()\n best_score = float('inf')\n\n def rescale_sequence(seq):\n \"\"\"No rescaling, keep original sequence\"\"\"\n return [max(0.0, x) for x in seq]\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using improved line search and constraint selection.\"\"\"\n n = len(sequence)\n if n == 0:\n return None\n\n sum_sequence = np.sum(sequence)\n conv = np.convolve(sequence, sequence)\n max_b_val = np.max(conv)\n\n # Select most impactful constraints\n top_indices = np.argsort(conv)[::-1][:int(len(conv) * 0.3)]\n if not top_indices.size:\n top_indices = np.arange(2 * n - 1)\n\n # Solve LP with selected constraints\n g_fun = solve_convolution_lp(sequence, max_b_val, top_indices)\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 # Perform binary search on t to find optimal combination\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 # Binary search bounds\n low, high = 0.0, 1.0\n best_t = 0.0\n best_score = float('inf')\n for _ in range(50): # 50 iterations of binary search\n mid = (low + high) / 2\n curr_score = objective(mid)\n if curr_score < best_score:\n best_score = curr_score\n best_t = mid\n if curr_score < evaluate_sequence(sequence):\n high = mid\n else:\n low = mid\n\n if best_score < float('inf'):\n return [(1 - best_t) * x + best_t * y for x, y in zip(sequence, g_fun)]\n return None\n\n def solve_convolution_lp(f_sequence, rhs, tight_positions):\n \"\"\"Solves LP with enhanced constraint selection.\"\"\"\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)\n objective = cp.Maximize(cp.sum(g))\n\n # Constraints: ensure max(convolution) <= rhs\n constraints = []\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.SCS, verbose=False)\n if problem.status == cp.OPTIMAL:\n return g.value\n except:\n return None\n return None\n\n def perturb_sequence(seq):\n \"\"\"Perturb selected elements while maintaining sum normalization.\"\"\"\n n = len(seq)\n conv = np.convolve(seq, seq)\n top_indices = np.argsort(conv)[-30:]\n indices = np.random.choice(top_indices, size=min(30, len(top_indices)), replace=False)\n indices = [min(idx, n-1) for idx in indices]\n for idx in indices:\n scale = 0.05 * np.std(seq)\n new_val = max(0.0, seq[idx] + np.random.exponential(scale) - scale * 0.3)\n seq = [new_val if i == idx else seq[i] for i in range(n)]\n return seq\n\n def random_walk(seq):\n \"\"\"Add small random perturbations to all elements.\"\"\"\n n = len(seq)\n scale = 0.05 * np.std(seq)\n new_seq = [max(0.0, x + np.random.normal(0, scale)) for x in 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 current_sequence = perturb_sequence(current_sequence)\n else:\n current_sequence = h_function\n\n # Occasionally try random walk for exploration\n if np.random.rand() < 0.1:\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": 265.6219766661525,
"env/all/time/policy/min": 106.65047073364258,
"env/all/time/policy/max": 348.4655876159668,
"env/all/time/env_step": 2256.6698965686373,
"env/all/time/env_step/min": 0.006376028060913086,
"env/all/time/env_step/max": 4969.078485250473,
"env/all/time/reward_compute": 1.032371073961258e-06,
"env/all/time/reward_compute/min": 1.2293457984924316e-07,
"env/all/time/reward_compute/max": 3.080815076828003e-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": 0.03340604901313782,
"advantage/min": -1.0,
"advantage/max": 20.42847442626953,
"time/assemble_training_data": 8.612946510314941,
"time/kl_vs_base": 86.92946434020996,
"kl_policy_base": 0.0008762289653532207,
"time/train": 516.5952305793762,
"time/save_checkpoint": 14.832954406738281,
"time/total": 5943.6048579216
}
[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