Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step8
ttt-discover-v3-qwen3-8b-ac1-step8 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 7 (0-indexed). Strict upstream eval parity: 1100s hard k…
Downloads · 30 days
0
Access
Public
Updated Jul 9, 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 7 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 7,
"progress/batch": 7,
"optim/lr": 4e-05,
"progress/done_frac": 0.16,
"puct/buffer_size": 120,
"puct/sampled_size": 8,
"puct/T": 3584,
"puct/scale_last": 0.3984176677464917,
"puct/buffer_value/mean": -1.5599740057031894,
"puct/buffer_value/std": 0.13573406913487582,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5072053730475787,
"puct/buffer_timestep/mean": 2.7333333333333334,
"puct/buffer_timestep/std": 2.174600857373345,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 6.0,
"puct/buffer_construction_len/mean": 1362.7583333333334,
"puct/buffer_construction_len/std": 1181.8780182110822,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5072679622672172,
"puct/sampled_value/std": 3.73754006684519e-05,
"puct/sampled_value/min": -1.507306554705767,
"puct/sampled_value/max": -1.5072053730475787,
"puct/sampled_timestep/mean": 6.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 6.0,
"puct/sampled_timestep/max": 6.0,
"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": 5823.462568998337,
"env/all/ac_tokens_per_turn": 8579.087890625,
"env/all/ob_tokens_per_turn": 3267.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4392493,
"env/all/total_ob_tokens": 1673024,
"env/all/time/sampling_mean": 314.8483235761523,
"env/all/time/sampling_max": 402.0293309688568,
"env/all/time/env_step_mean": 2777.3003276418895,
"env/all/time/env_step_max": 5416.754773855209,
"env/all/reward/mean": 0.5841408910981827,
"env/all/reward/max": 0.663487262938591,
"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": 0.5841408910981827,
"env/all/correctness": 0.900390625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 3.131454960202686,
"env/all/raw_score/min": 1.5071879284255112,
"env/all/raw_score/max": 537.6478181854093,
"env/all/initial_raw_score": -1.5072679622672172,
"env/all/initial_raw_score/min": -1.507306554705767,
"env/all/initial_raw_score/max": -1.5072053730475787,
"env/all/msg": "Success; raw_score=1.5072050236684411",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy import optimize\nimport random\n\nlinprog = optimize.linprog\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers that minimizes the evaluation score \n using a hybrid of linear programming with adaptive line search and targeted perturbations.\n Focuses on improving initial sequences and enhancing line search for better convergence.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n\n # Use the last best construction if available\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 # Generate a better initial sequence: random uniform sequence normalized\n n = 1000\n random_seq = np.random.uniform(0.0, 1.0, n)\n random_seq *= np.sqrt(2 * n) / np.sum(random_seq)\n best_sequence = random_seq.tolist()\n \n curr_sequence = [float(x) for x in best_sequence]\n best_score = evaluate_sequence(curr_sequence)\n\n while time.time() < deadline:\n try:\n h_function = get_good_direction_to_move_into(curr_sequence)\n if h_function is not None:\n curr_sequence = h_function\n else:\n # Apply aggressive targeted perturbation\n idx = perturb_targeted(curr_sequence)\n # Larger perturbation to escape local minima\n curr_sequence = perturb_sequence(curr_sequence, idx, 20.0)\n except Exception:\n print(\"Evaluation error, 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 # Final cleanup to ensure non-negative values\n return [float(max(0.0, x)) for x in best_sequence]\n\ndef get_good_direction_to_move_into(sequence):\n \"\"\"Returns a better sequence using LP with enhanced 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 LP\n normalized_sequence = [x * np.sqrt(2 * n) / sum_sequence for x in sequence]\n rhs = np.max(np.convolve(normalized_sequence, normalized_sequence))\n\n # Solve LP\n g_fun = solve_convolution_lp(normalized_sequence, 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_fun = [x * np.sqrt(2 * n) / sum_g for x in g_fun]\n # Perform more refined line search\n def score_func(t):\n new_seq = [(1 - t) * x + t * y for x, y in zip(sequence, normalized_g_fun)]\n try:\n new_score = evaluate_sequence(new_seq)\n return new_score\n except:\n return np.inf\n # Golden section search to find optimal t\n a, b = 0.0, 1.0\n phi = (np.sqrt(5) - 1) / 2\n for _ in range(100):\n c = b - phi * (b - a)\n d = a + phi * (b - a)\n if score_func(c) < score_func(d):\n b = d\n else:\n a = c\n t_opt = (a + b) / 2\n new_sequence = [(1 - t_opt) * x + t_opt * y for x, y in zip(sequence, normalized_g_fun)]\n return new_sequence\n \n # If LP fails, find max_conv_idx\n conv = np.convolve(normalized_sequence, normalized_sequence)\n max_conv_idx = np.argmax(conv)\n max_conv_val = conv[max_conv_idx]\n # Identify contributing elements\n perturb_elements = set()\n for i in range(n):\n j = max_conv_idx - i\n if 0 <= j < n:\n perturb_elements.add(i)\n perturb_elements.add(j)\n if perturb_elements:\n idx = perturb_targeted(normalized_sequence, max_conv_idx)\n new_seq = normalized_sequence.copy()\n new_seq[idx] = max(0.0, new_seq[idx] - np.random.randn() * 3.0)\n return new_seq\n else:\n idx = np.random.randint(0, len(normalized_sequence))\n new_seq = normalized_sequence.copy()\n new_seq[idx] = max(0.0, new_seq[idx] - np.random.randn() * 2.0)\n return new_seq\n\ndef perturb_targeted(sequence, max_conv_idx):\n \"\"\"Targeted perturbation of elements contributing to max_conv_idx\"\"\"\n n = len(sequence)\n contributions = []\n for i in range(n):\n j = max_conv_idx - i\n if 0 <= j < n:\n contributions.append((i, abs(sequence[i] * sequence[j])))\n if contributions:\n sorted_contributions = sorted(contributions, key=lambda x: x[1], reverse=True)\n return sorted_contributions[0][0]\n else:\n return np.random.randint(0, n)\n\ndef perturb_sequence(sequence, idx, perturb_amount):\n \"\"\"Perturb the sequence at given index with larger amount\"\"\"\n new_seq = sequence.copy()\n new_seq[idx] = max(0.0, new_seq[idx] - np.random.randn() * perturb_amount)\n new_seq[idx] = min(1000.0, new_seq[idx])\n return new_seq\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)\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\n a_ub_nonneg = -np.eye(n)\n b_ub_nonneg = np.zeros(n)\n a_ub = np.vstack([a_ub, a_ub_nonneg])\n b_ub = np.hstack([b_ub, b_ub_nonneg])\n\n # Use resilient LP solver\n result = 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\": 30.0, # Reduced time for faster exploration\n \"disp\": False,\n },\n )\n\n if result.success:\n return result.x\n return None\n```",
"env/all/time/policy": 314.8483235761523,
"env/all/time/policy/min": 143.38468980789185,
"env/all/time/policy/max": 402.0293309688568,
"env/all/time/env_step": 2777.3003276418895,
"env/all/time/env_step/min": 0.006407022476196289,
"env/all/time/env_step/max": 5416.754773855209,
"env/all/time/reward_compute": 3.390014171600342e-07,
"env/all/time/reward_compute/min": 2.3096799850463867e-07,
"env/all/time/reward_compute/max": 7.264316082000732e-07,
"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": 0.03156890720129013,
"advantage/min": -1.0,
"advantage/max": 23.451736450195312,
"time/assemble_training_data": 10.100822925567627,
"time/kl_vs_base": 91.52989554405212,
"kl_policy_base": 0.0006722698453813791,
"time/train": 567.9420261383057,
"time/save_checkpoint": 35.07109999656677,
"time/total": 6530.11212015152
}
[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