Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step17
ttt-discover-v3-qwen3-8b-ac1-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": 0.39924765732831347,
"puct/buffer_value/mean": -1.5311399762911762,
"puct/buffer_value/std": 0.09524307699881022,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.506375383465757,
"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": 1164.8901515151515,
"puct/buffer_construction_len/std": 817.0384384488941,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5063766849380247,
"puct/sampled_value/std": 1.2051593746580878e-06,
"puct/sampled_value/min": -1.5063789030684447,
"puct/sampled_value/max": -1.506375383465757,
"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": 5852.22451710701,
"env/all/ac_tokens_per_turn": 8577.935546875,
"env/all/ob_tokens_per_turn": 3114.5,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4391903,
"env/all/total_ob_tokens": 1594624,
"env/all/time/sampling_mean": 303.8601933433674,
"env/all/time/sampling_max": 397.3574478626251,
"env/all/time/env_step_mean": 2868.711318344809,
"env/all/time/env_step_max": 5452.137911081314,
"env/all/reward/mean": 0.6310106626052941,
"env/all/reward/max": 0.6638509322715221,
"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": 0.6310106626052941,
"env/all/correctness": 0.962890625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.548814464849001,
"env/all/raw_score/min": 1.5063622641000978,
"env/all/raw_score/max": 11.413581485426553,
"env/all/initial_raw_score": -1.5063766849380251,
"env/all/initial_raw_score/min": -1.5063789030684447,
"env/all/initial_raw_score/max": -1.506375383465757,
"env/all/msg": "Success; raw_score=1.506375383465757",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy import optimize\nimport cvxpy as cp\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score \n using enhanced gradient-based direction exploration and adaptive solver configuration.\n Focuses on reducing the maximum convolution value via structured perturbation and\n direct LP-based optimization.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Initialize using the best known sequence or a uniform initialization\n n = 1000\n best_sequence = None\n if 'height_sequence_1' in globals() and isinstance(height_sequence_1, (list, np.ndarray)):\n best_sequence = list(height_sequence_1)\n else:\n # Uniform initialization, normalized to maintain sum_a = sqrt(2n)\n uniform_seq = np.ones(n) * np.sqrt(2 * n) / n\n best_sequence = uniform_seq.tolist()\n \n curr_sequence = [float(x) for x in best_sequence]\n best_score = evaluate_sequence(curr_sequence)\n\n # Parameters for perturbation and blending\n perturb_rate = 0.1 # Increased perturbation rate for more exploration\n blend_factor = 0.7 # Adjusted blend factor\n min_perturb = 0.005\n max_iter = 500 # Limit to prevent infinite loops\n\n for iter_count in range(max_iter):\n if time.time() >= deadline:\n break\n \n try:\n # Compute current convolution and max_b\n conv = np.convolve(curr_sequence, curr_sequence)\n max_b = np.max(conv)\n sum_a = np.sum(curr_sequence)\n if sum_a < 0.01:\n raise ValueError(\"Sum too small\")\n\n # Try direct LP optimization with tightened constraints\n try:\n # Solve a convex program to minimize maximum convolution value\n a = cp.Variable(n, nonneg=True)\n M = cp.Variable(nonneg=True)\n sum_a_expr = cp.sum(a)\n constraints = [sum_a_expr == np.sqrt(2 * n)]\n for k in range(2 * n - 1):\n val = 0.0\n for i in range(n):\n j = k - i\n if 0 <= j < n:\n val += a[i] * a[j]\n constraints.append(val <= M)\n objective = cp.Minimize(M)\n prob = cp.Problem(objective, constraints)\n prob.solve()\n if prob.status == 'OPTIMAL':\n if M.value < max_b:\n curr_sequence = [float(x) for x in a.value]\n best_score = evaluate_sequence(curr_sequence)\n print(f\"New best score via LP: {best_score}\")\n continue\n except:\n pass # Fallback if LP fails\n\n # Compute LP solution as a fallback\n normalized_seq = [x / sum_a * np.sqrt(2 * n) for x in curr_sequence]\n rhs = max_b\n g_fun = solve_convolution_lp(normalized_seq, rhs)\n if g_fun is not None and np.sum(g_fun) > 0.0:\n sum_g = np.sum(g_fun)\n normalized_g = [x / sum_g * np.sqrt(2 * n) for x in g_fun]\n \n # Evaluate the LP solution directly\n new_score = evaluate_sequence(normalized_g)\n if new_score < best_score:\n curr_sequence = normalized_g\n curr_score = new_score\n best_score = curr_score\n print(f\"New best score: {best_score}\")\n continue\n else:\n # Proceed with blending\n def objective(t):\n new_seq = [(1 - t) * x + t * y for x, y in zip(curr_sequence, normalized_g)]\n return evaluate_sequence(new_seq)\n \n result = optimize.minimize_scalar(objective, bounds=(0.0, 1.0), method='bounded', tol=1e-5)\n best_t = result.x\n new_seq = [(1 - best_t) * x + best_t * y for x, y in zip(curr_sequence, normalized_g)]\n curr_sequence = new_seq\n else:\n # Fallback to targeted perturbation\n conv = np.convolve(curr_sequence, curr_sequence)\n max_idx = np.argmax(conv)\n contributions = []\n for i in range(n):\n j = max_idx - i\n if 0 <= j < n:\n contributions.append((i, abs(curr_sequence[i] * curr_sequence[j])))\n if contributions:\n sorted_contributions = sorted(contributions, key=lambda x: x[1], reverse=True)\n idxs = [sorted_contributions[i][0] for i in range(min(5, len(sorted_contributions)))]\n curr_sequence = perturb_sequence(curr_sequence, idxs, perturb_rate)\n else:\n idx = np.random.randint(0, n)\n curr_sequence = perturb_sequence(curr_sequence, [idx], perturb_rate)\n\n except Exception as e:\n print(f\"Error during optimization: {e}. Skipping update.\")\n continue\n\n try:\n curr_score = evaluate_sequence(curr_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = curr_sequence.copy()\n print(f\"New best score: {best_score}\")\n except Exception:\n print(\"Evaluation error, skipping update\")\n\n return [float(max(0.0, x)) for x in best_sequence]\n\ndef solve_convolution_lp(f_sequence, rhs):\n \"\"\"Solves LP to maximize sum(b) s.t. conv(f, b) <= rhs, b >= 0\"\"\"\n n = len(f_sequence)\n if n == 0:\n return None\n\n c = -np.ones(n) # Maximize sum(b) via -1 coefficients\n a_ub = []\n b_ub = []\n\n for k in range(2 * n - 1):\n row = np.zeros(n)\n for i in range(n):\n j = k - i\n if 0 <= j < n:\n row[j] = f_sequence[i]\n a_ub.append(row)\n b_ub.append(rhs)\n\n # Non-negativity constraints (already handled by bounds)\n try:\n result = optimize.linprog(\n c,\n A_ub=a_ub,\n b_ub=b_ub,\n bounds=(0.0, 1000.0),\n method='highs',\n options={\n \"time_limit\": 300.0, # Increased time limit for faster iterations\n \"disp\": False,\n },\n )\n if result.success:\n return result.x\n except:\n pass\n return None\n\ndef perturb_sequence(sequence, idxs, perturb_amount):\n \"\"\"Perturb multiple indices in sequence with controlled amounts.\"\"\"\n new_seq = sequence.copy()\n for idx in idxs:\n new_seq[idx] = max(0.0, new_seq[idx] - perturb_amount * new_seq[idx])\n new_seq[idx] = min(1000.0, new_seq[idx])\n return new_seq\n```",
"env/all/time/policy": 303.8601933433674,
"env/all/time/policy/min": 97.78522968292236,
"env/all/time/policy/max": 397.3574478626251,
"env/all/time/env_step": 2868.711318344809,
"env/all/time/env_step/min": 0.009800195693969727,
"env/all/time/env_step/max": 5452.137911081314,
"env/all/time/reward_compute": 5.885958671569824e-07,
"env/all/time/reward_compute/min": 2.2724270820617676e-07,
"env/all/time/reward_compute/max": 1.8477439880371094e-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": 0.011791655793786049,
"advantage/min": -1.0,
"advantage/max": 10.428145408630371,
"time/assemble_training_data": 8.205443620681763,
"time/kl_vs_base": 96.11744999885559,
"kl_policy_base": 0.0008120968122966588,
"time/train": 559.9850475788116,
"time/save_checkpoint": 33.46625208854675,
"time/total": 6552.008772850037
}
[2026-07-09T06:24:18+00:00] job=1812628 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:30:32+00:00] job=1813127 node=node-22 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:07:08+00:00] job=1813128 node=node-6 ngpu=6 ntrain=2 replicas=4 flash_attn=yes