Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step9
ttt-discover-v3-qwen3-8b-ac1-s2-step9 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 8 (0-indexed). Strict upstream eval parity: 1100s hard k…
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 8 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 8,
"progress/batch": 8,
"optim/lr": 4e-05,
"progress/done_frac": 0.18,
"puct/buffer_size": 136,
"puct/sampled_size": 8,
"puct/T": 4096,
"puct/scale_last": 0.49195678802386644,
"puct/buffer_value/mean": -1.5686778518712559,
"puct/buffer_value/std": 0.14437884457281733,
"puct/buffer_value/min": -2.000000000000007,
"puct/buffer_value/max": -1.5063036919506345,
"puct/buffer_timestep/mean": 3.235294117647059,
"puct/buffer_timestep/std": 2.462170533700747,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 7.0,
"puct/buffer_construction_len/mean": 1441.1029411764705,
"puct/buffer_construction_len/std": 1360.9573310909395,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5067172404138414,
"puct/sampled_value/std": 0.00044627752980154104,
"puct/sampled_value/min": -1.507319898567202,
"puct/sampled_value/max": -1.5063036919506345,
"puct/sampled_timestep/mean": 7.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 7.0,
"puct/sampled_timestep/max": 7.0,
"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": 5270.708781242371,
"env/all/ac_tokens_per_turn": 8088.037109375,
"env/all/ob_tokens_per_turn": 2994.875,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4141075,
"env/all/total_ob_tokens": 1533376,
"env/all/time/sampling_mean": 280.3904445078224,
"env/all/time/sampling_max": 368.19845962524414,
"env/all/time/env_step_mean": 2291.6180606423877,
"env/all/time/env_step_max": 4919.026139259338,
"env/all/reward/mean": 0.49376736442555696,
"env/all/reward/max": 0.6638767459078934,
"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": 0.49376736442555696,
"env/all/correctness": 0.76953125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.619619332526876,
"env/all/raw_score/min": 1.5063036919506338,
"env/all/raw_score/max": 18.054986317711485,
"env/all/initial_raw_score": -1.5067172404138416,
"env/all/initial_raw_score/min": -1.507319898567202,
"env/all/initial_raw_score/max": -1.5063036919506345,
"env/all/msg": "Success; raw_score=1.5063036919506358",
"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 np.isscalar(prev):\n prev = []\n for _ in range(1000):\n prev.append(prev)\n if prev is not None and len(prev) > 0:\n best_sequence = list(np.asarray(prev, dtype=float))\n else:\n n = 1000\n base = 0.2 * np.ones(n)\n best_sequence = [max(0.0, base[i]) for i in range(n)]\n \n current_sequence = best_sequence.copy()\n best_score = float('inf')\n\n def get_good_direction_to_move_into(sequence):\n \"\"\"Computes a better sequence using advanced LP with adaptive constraints and optimized step size.\"\"\"\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 normalized_sequence = [x * np.sqrt(2 * n) / sum_sequence for x in sequence]\n conv = np.convolve(normalized_sequence, normalized_sequence)\n max_b = np.max(conv)\n\n # Generate tight positions with increased tolerance\n current_conv = conv\n max_b_val = max_b\n tolerance = 1e-4 # Increased tolerance to include more positions\n tight_positions = np.where(np.isclose(current_conv, max_b_val, atol=tolerance))[0]\n if len(tight_positions) < 5:\n # Fallback to positions 0 to 2n-1\n tight_positions = np.arange(2 * n - 1)\n if not tight_positions.size:\n tight_positions = np.arange(2 * n - 1)\n\n # Solve LP with extended time limit\n g_fun = solve_convolution_lp(normalized_sequence, max_b, 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 normalized_g = [x * np.sqrt(2 * n) / sum_g for x in g_fun]\n\n # Adaptive line search with refined t values\n best_t = 0.001\n best_score = float('inf')\n # Use a binary search on t to find optimal step\n low, high = 0.0, 1.0\n for _ in range(50): # increased iterations\n mid = (low + high) / 2\n try:\n new_sequence = [(1 - mid) * x + mid * y for x, y in zip(sequence, normalized_g)]\n curr_score = evaluate_sequence(new_sequence)\n if curr_score < best_score:\n best_t = mid\n best_score = curr_score\n except Exception:\n continue\n # Adjust bounds based on improvement\n if curr_score < best_score:\n high = mid\n else:\n low = mid\n\n return [(1 - best_t) * x + best_t * y for x, y in zip(sequence, normalized_g)]\n\n def solve_convolution_lp(f_sequence, rhs, tight_positions):\n \"\"\"Solves LP with tight positions using cvxpy with GLOP solver.\"\"\"\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), equivalent to minimize -sum(g_j)\n objective = cp.Minimize(-cp.sum(g))\n\n # Constraints\n constraints = []\n for k in tight_positions:\n # Coefficients for the constraint: sum_{j=0}^{n-1} f[i] * g[j] <= rhs for appropriate i\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\n problem = cp.Problem(objective, constraints)\n try:\n problem.solve(solver=cp.GLPK, 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 \"\"\"Perturbation with multiple strategies to avoid local optima.\"\"\"\n conv = np.convolve(seq, seq)\n max_conv_pos = np.argmax(conv)\n # Perturb one element with higher variance\n indices = np.random.choice(len(seq), size=1, replace=False)\n for idx in indices:\n new_val = max(0.0, seq[idx] + np.random.normal(0, 0.15))\n seq = [new_val if i == idx else seq[i] for i in range(len(seq))]\n return seq\n\n while time.time() < deadline:\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 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": 280.3904445078224,
"env/all/time/policy/min": 112.9582347869873,
"env/all/time/policy/max": 368.19845962524414,
"env/all/time/env_step": 2291.6180606423877,
"env/all/time/env_step/min": 0.00561213493347168,
"env/all/time/env_step/max": 4919.026139259338,
"env/all/time/reward_compute": 3.7159770727157593e-07,
"env/all/time/reward_compute/min": 2.3096799850463867e-07,
"env/all/time/reward_compute/max": 1.0170042514801025e-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": 0.018856126815080643,
"advantage/min": -1.0,
"advantage/max": 10.54602336883545,
"time/assemble_training_data": 5.540218114852905,
"time/kl_vs_base": 93.20460057258606,
"kl_policy_base": 0.0007354775443673134,
"time/train": 525.7338175773621,
"time/save_checkpoint": 17.851184129714966,
"time/total": 5914.644565343857
}
[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