Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step15
ttt-discover-v3-qwen3-8b-ac1-step15 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 14 (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 14 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 14,
"progress/batch": 14,
"optim/lr": 4e-05,
"progress/done_frac": 0.3,
"puct/buffer_size": 232,
"puct/sampled_size": 8,
"puct/T": 7168,
"puct/scale_last": 0.39923677498222543,
"puct/buffer_value/mean": -1.5344104265207177,
"puct/buffer_value/std": 0.10114219462348664,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.506386265811845,
"puct/buffer_timestep/mean": 6.241379310344827,
"puct/buffer_timestep/std": 4.190754786874301,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 13.0,
"puct/buffer_construction_len/mean": 1187.6336206896551,
"puct/buffer_construction_len/std": 869.1147692597976,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5064400947669205,
"puct/sampled_value/std": 4.9116637451801e-05,
"puct/sampled_value/min": -1.5064970927205832,
"puct/sampled_value/max": -1.506386265811845,
"puct/sampled_timestep/mean": 13.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 13.0,
"puct/sampled_timestep/max": 13.0,
"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": 6165.917045116425,
"env/all/ac_tokens_per_turn": 8561.404296875,
"env/all/ob_tokens_per_turn": 3474.5,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4383439,
"env/all/total_ob_tokens": 1778944,
"env/all/time/sampling_mean": 313.7926362338476,
"env/all/time/sampling_max": 404.19773387908936,
"env/all/time/env_step_mean": 2825.1796453464776,
"env/all/time/env_step_max": 5752.865364789963,
"env/all/reward/mean": 0.5801304011384645,
"env/all/reward/max": 0.6638449057567727,
"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": 0.5801304011384645,
"env/all/correctness": 0.912109375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 2.135077791646538,
"env/all/raw_score/min": 1.5063759391533882,
"env/all/raw_score/max": 180.73813334099864,
"env/all/initial_raw_score": -1.5064400947669203,
"env/all/initial_raw_score/min": -1.5064970927205832,
"env/all/initial_raw_score/max": -1.506386265811845,
"env/all/msg": "Success; raw_score=1.5064586332331835",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy import optimize\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 LP-based directions, adaptive blending, and aggressive perturbations.\n Focuses on improving solver efficiency, blending strategies, and perturbation dynamics.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Initialize using the best known sequence or a structured Gaussian + uniform\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 # Gaussian-initialized sequence, normalized and clamped\n random_seq = np.random.normal(0.5, 0.15, n)\n random_seq = np.maximum(0.0, np.clip(random_seq, 0, 1000.0))\n random_seq *= np.sqrt(2 * n) / np.sum(random_seq)\n best_sequence = random_seq.tolist()\n \n # Start with a combination of the Gaussian and uniform\n uniform_seq = np.ones(n) * np.sqrt(2 * n) / n\n best_sequence = [float(x) for x in best_sequence]\n # Start with a combination of both initializations (weight 0.5 each)\n best_sequence = [x * 0.5 + uniform_seq[i] * 0.5 for i, x in enumerate(best_sequence)]\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 = 0.25 # Increased from 0.1\n blend_factor = 0.35 # Adjusted blend factor\n min_perturb = 0.01\n # Parameters for aggressive perturbation\n max_perturb = 0.4\n aggressive_perturb_rate = 0.1\n aggressive_perturb_steps = 2\n\n while time.time() < deadline:\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 # Normalize for LP\n normalized_seq = [x / sum_a * np.sqrt(2 * n) for x in curr_sequence]\n rhs = max_b\n\n # Solve LP to maximize sum(b) under convolution constraint\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 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 with aggressive steps\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)))] # Perturb more elements\n # Apply aggressive perturbation with multiple steps\n for _ in range(aggressive_perturb_steps):\n curr_sequence = perturb_sequence(curr_sequence, idxs, aggressive_perturb_rate)\n else:\n idx = np.random.randint(0, n)\n curr_sequence = perturb_sequence(curr_sequence, [idx], aggressive_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\": 120.0, # Extended time limit for LP\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": 313.7926362338476,
"env/all/time/policy/min": 139.18307328224182,
"env/all/time/policy/max": 404.19773387908936,
"env/all/time/env_step": 2825.1796453464776,
"env/all/time/env_step/min": 0.008219003677368164,
"env/all/time/env_step/max": 5752.865364789963,
"env/all/time/reward_compute": 9.587965905666351e-07,
"env/all/time/reward_compute/min": 2.3096799850463867e-07,
"env/all/time/reward_compute/max": 2.9169023036956787e-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": 0.02859635278582573,
"advantage/min": -1.0,
"advantage/max": 24.416460037231445,
"time/assemble_training_data": 10.338034868240356,
"time/kl_vs_base": 100.38881802558899,
"kl_policy_base": 0.0007811358664184809,
"time/train": 579.3532009124756,
"time/save_checkpoint": 35.78655004501343,
"time/total": 6894.131791353226
}
[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