Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac2-s2-step12
ttt-discover-v3-qwen3-8b-ac2-s2-step12 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 ac2. Checkpoint saved after training step 11 (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 ac2. Checkpoint saved
after training step 11 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 11,
"progress/batch": 11,
"optim/lr": 4e-05,
"progress/done_frac": 0.24,
"puct/buffer_size": 181,
"puct/sampled_size": 8,
"puct/T": 5632,
"puct/scale_last": 0.18873072556780646,
"puct/buffer_value/mean": 0.9150509867667729,
"puct/buffer_value/std": 0.059506889630791204,
"puct/buffer_value/min": 0.6666666666666636,
"puct/buffer_value/max": 0.9451865858077976,
"puct/buffer_timestep/mean": 4.767955801104972,
"puct/buffer_timestep/std": 3.3409018397760257,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 10.0,
"puct/buffer_construction_len/mean": 3291.6187845303866,
"puct/buffer_construction_len/std": 2896.574069631676,
"puct/buffer_construction_len/min": 1024.0,
"puct/buffer_construction_len/max": 32768.0,
"puct/sampled_value/mean": 0.944043944243937,
"puct/sampled_value/std": 0.000632835817209034,
"puct/sampled_value/min": 0.9432134587891141,
"puct/sampled_value/max": 0.9451865858077976,
"puct/sampled_timestep/mean": 10.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 10.0,
"puct/sampled_timestep/max": 10.0,
"puct/sampled_construction_len/mean": 4096.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 4096.0,
"puct/sampled_construction_len/max": 4096.0,
"time/sampling": 5720.409188270569,
"env/all/ac_tokens_per_turn": 8550.5546875,
"env/all/ob_tokens_per_turn": 4215.875,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4377884,
"env/all/total_ob_tokens": 2158528,
"env/all/time/sampling_mean": 326.21510665211827,
"env/all/time/sampling_max": 548.1280097961426,
"env/all/time/env_step_mean": 2489.2982726749033,
"env/all/time/env_step_max": 5279.992500782013,
"env/all/reward/mean": 0.1560041383976076,
"env/all/reward/max": 0.9459495264324065,
"env/all/reward/min": 0.0,
"env/all/format": 0.994140625,
"env/all/format/min": 0.0,
"env/all/format/max": 1.0,
"env/all/reward": 0.1560041383976076,
"env/all/correctness": 0.17578125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.8874902095508345,
"env/all/raw_score/min": 0.3175793859770932,
"env/all/raw_score/max": 0.9459495264324065,
"env/all/initial_raw_score": 0.9440439442439368,
"env/all/initial_raw_score/min": 0.9432134587891141,
"env/all/initial_raw_score/max": 0.9451865858077976,
"env/all/msg": "Evaluation timed out after 1100 seconds.",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom typing import Tuple, List\nimport random\nimport time\n\n# Re-define helper functions for evaluating and differentiating the loss\ndef _simpson_l2sq(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n m = conv.size\n if m == 0:\n return 0.0, np.zeros_like(conv)\n dx = 1.0 / (m + 1)\n y = np.zeros(m + 2, dtype=conv.dtype)\n y[0] = 0.0\n y[1:-1] = conv\n y[-1] = 0.0\n lhs = y[:-1]\n rhs = y[1:]\n l2_sq = (dx / 3.0) * np.sum(lhs * lhs + lhs * rhs + rhs * rhs)\n grad_y = (dx / 3.0) * (4.0 * y + np.roll(y, 1) + np.roll(y, -1))\n grad_conv = grad_y[1:-1]\n return float(l2_sq), grad_conv\n\ndef _l1(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n m = conv.size\n dx = 1.0 / (m + 1) if m > 0 else 1.0\n val = dx * float(np.sum(conv)) if m > 0 else 0.0\n grad = np.full_like(conv, dx)\n return val, grad\n\ndef _linf(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n if conv.size == 0:\n return 0.0, np.zeros_like(conv)\n m = float(np.max(conv))\n if m <= 0.0:\n return m, np.zeros_like(conv)\n mask = conv == m\n count = int(mask.sum())\n if count == 0:\n return m, np.zeros_like(conv)\n grad = mask.astype(conv.dtype) / count\n return m, grad\n\ndef _objective_and_grad_conv(conv: np.ndarray) -> Tuple[float, np.ndarray]:\n l2_sq, g_l2 = _simpson_l2sq(conv)\n l1, g_l1 = _l1(conv)\n linf, g_linf = _linf(conv)\n if l1 <= 0.0 or linf <= 0.0:\n return 0.0, np.zeros_like(conv)\n denom = l1 * linf\n c_value = l2_sq / denom\n num_grad = g_l2 * denom - l2_sq * (g_l1 * linf + l1 * g_linf)\n g_conv = num_grad / (denom * denom)\n return float(c_value), g_conv\n\ndef _grad_h_from_conv_grad(h: np.ndarray, g_conv: np.ndarray) -> np.ndarray:\n h_rev = h[::-1]\n g_h = np.convolve(g_conv, h_rev, mode=\"valid\")\n return 2.0 * g_h\n\ndef _upsample_1d(h: np.ndarray) -> np.ndarray:\n n = h.shape[0]\n x_old = np.linspace(-0.5, 0.5, n)\n x_new = np.linspace(-0.5, 0.5, 2 * n)\n return np.interp(x_new, x_old, h)\n\ndef _single_candidate_finetune(h0: np.ndarray, lr=5e-3, steps=300_000) -> Tuple[np.ndarray, float]:\n h = h0.astype(np.float32).copy()\n opt = _Adam(h.shape, lr=lr, dtype=h.dtype)\n last_c = 0.0\n for _ in range(steps):\n h_clip = np.clip(h, 0.0, None)\n conv = np.convolve(h_clip, h_clip, mode=\"full\")\n c_val, g_conv = _objective_and_grad_conv(conv)\n g_h = _grad_h_from_conv_grad(h_clip, g_conv)\n h = np.clip(opt.step(h, g_h.astype(h.dtype)), 0.0, None)\n last_c = c_val\n return h, float(last_c)\n\nclass _Adam:\n def __init__(self, shape, lr=3e-2, beta1=0.9, beta2=0.999, eps=1e-8, dtype=np.float32):\n self.m = np.zeros(shape, dtype=dtype)\n self.v = np.zeros(shape, dtype=dtype)\n self.t = 0\n self.lr = lr\n self.b1 = beta1\n self.b2 = beta2\n self.eps = eps\n\n def step(self, params, grad):\n self.t += 1\n self.m = self.b1 * self.m + (1 - self.b1) * grad\n self.v = self.b2 * self.v + (1 - self.b2) * (grad * grad)\n m_hat = self.m / (1 - self.b1 ** self.t)\n v_hat = self.v / (1 - self.b2 ** self.t)\n return params + self.lr * m_hat / (np.sqrt(v_hat) + self.eps)\n\ndef construct_function():\n \"\"\"\n Optimizes a sequence of non-negative heights for maximizing the evaluation score,\n using a combination of adaptive gradient ascent, multi-scale refinement, and more diverse initial candidates.\n The strategy explores structured and random initial sequences and fine-tunes them with increased refinement steps.\n \"\"\"\n np.random.seed(0)\n \n # Load previous best sequence\n height_sequence_1 = globals().get(\"height_sequence_1\", None)\n initial_n = 512 # Starting sequence length\n initial_candidates = []\n \n if height_sequence_1 is not None:\n # Use previous best and upscale\n initial_h = np.array(height_sequence_1, dtype=np.float32)\n initial_n = min(len(initial_h), 2048)\n x_old = np.linspace(-0.5, 0.5, len(initial_h))\n x_new = np.linspace(-0.5, 0.5, initial_n)\n h = np.interp(x_new, x_old, initial_h)\n initial_candidates.append(h)\n else:\n # Generate diverse initial sequences\n # Random\n initial_candidates.append(np.random.uniform(0.1, 1.0, initial_n))\n # Single peak\n single_peak = np.zeros(initial_n, dtype=np.float32)\n single_peak[initial_n // 2] = 1.0\n initial_candidates.append(single_peak)\n # Multi peaks\n multi_peak = np.zeros(initial_n, dtype=np.float32)\n for i in range(0, initial_n, initial_n // 4):\n multi_peak[i] = 1.0\n initial_candidates.append(multi_peak)\n # Flat\n flat_seq = np.ones(initial_n, dtype=np.float32) * 0.1\n initial_candidates.append(flat_seq)\n # Gaussian-like\n def generate_gaussian(n, peak=1.0, std_dev=0.3):\n x = np.linspace(-0.5, 0.5, n)\n y = np.exp(-((x) ** 2) / (2 * std_dev ** 2)) * peak\n return np.clip(y, 0.0, 1000.0)\n initial_candidates.append(generate_gaussian(initial_n))\n # Exponential decay\n def generate_exponential(n, initial_value=1.0, decay_rate=0.2):\n x = np.linspace(0, n-1, n)\n y = initial_value * np.exp(-decay_rate * x)\n return np.clip(y, 0.0, 1000.0)\n initial_candidates.append(generate_exponential(initial_n))\n # Symmetric\n def generate_symmetric(n):\n half = n // 2\n seq = np.zeros(n, dtype=np.float32)\n seq[:half] = np.random.uniform(0.1, 1.0, half)\n seq[half:] = seq[:half][::-1]\n return seq\n initial_candidates.append(generate_symmetric(initial_n))\n # Random peaks\n def generate_random_peaks(n, num_peaks=5):\n seq = np.zeros(n, dtype=np.float32)\n for _ in range(num_peaks):\n pos = np.random.randint(0, n)\n seq[pos] = np.random.uniform(0.5, 1.0)\n return seq\n initial_candidates.append(generate_random_peaks(initial_n))\n # Cosine-like\n def generate_cosine_like(n):\n x = np.linspace(0, n-1, n)\n seq = np.abs(np.cos(2 * np.pi * x / (n // 2))) * 0.8\n return np.clip(seq, 0.0, 1000.0)\n initial_candidates.append(generate_cosine_like(initial_n))\n \n best_score = -1.0\n best_heights = None\n \n # Try each initial sequence\n for candidate in initial_candidates:\n h = candidate.astype(np.float32).copy()\n h_sum = np.sum(h)\n if h_sum < 0.01:\n h = np.clip(h + (0.01 - h_sum) / initial_n, 0.0, 1000.0)\n \n # Gradient ascent\n learning_rate = 0.1\n noise_scale_initial = 0.3\n max_steps = 500_000\n upscale_steps = 100_000\n refine_steps = 300_000\n noise_decay = 0.97\n learning_rate_decay = 0.99\n \n start_time = time.time()\n \n # Gradient ascent\n for step in range(max_steps):\n clipped_h = np.clip(h, 0.0, None)\n conv = np.convolve(clipped_h, clipped_h, mode='full')\n obj_val, grad_conv = _objective_and_grad_conv(conv)\n grad_h = _grad_h_from_conv_grad(clipped_h, grad_conv)\n \n noise_scale = noise_scale_initial * (noise_decay ** step)\n learning_rate_current = learning_rate * (learning_rate_decay ** step)\n \n noise = noise_scale * np.random.normal(size=h.shape)\n h = np.clip(h + learning_rate_current * grad_h + noise, 0.0, 1000.0)\n \n h_sum = np.sum(h)\n if h_sum < 0.01:\n h = np.clip(h + (0.01 - h_sum) / h.shape[0], 0.0, 1000.0)\n \n remaining_time = 1000 - (time.time() - start_time)\n if remaining_time < 10:\n break\n \n # Upsample and refine\n h_up = _upsample_1d(h)\n h_up, _ = _single_candidate_finetune(h_up, lr=1e-2, steps=refine_steps)\n h_final = np.clip(h_up, 0.0, 1000.0)\n current_score = evaluate_sequence(h_final.tolist())\n \n if current_score > best_score:\n best_score = current_score\n best_heights = h_final.tolist()\n \n # Try multiple upscale steps\n if best_heights is not None:\n h = np.array(best_heights, dtype=np.float32)\n h_up = _upsample_1d(h)\n h_up, _ = _single_candidate_finetune(h_up, lr=1e-2, steps=200_000)\n h_final = np.clip(h_up, 0.0, 1000.0)\n current_score = evaluate_sequence(h_final.tolist())\n if current_score > best_score:\n best_score = current_score\n best_heights = h_final.tolist()\n \n return best_heights\n```",
"env/all/time/policy": 326.21510665211827,
"env/all/time/policy/min": 201.38559246063232,
"env/all/time/policy/max": 548.1280097961426,
"env/all/time/env_step": 2489.2982726749033,
"env/all/time/env_step/min": 0.007119655609130859,
"env/all/time/env_step/max": 5279.992500782013,
"env/all/time/reward_compute": 4.0140002965927124e-07,
"env/all/time/reward_compute/min": 2.2724270820617676e-07,
"env/all/time/reward_compute/max": 8.344650268554688e-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": 0.038163475692272186,
"advantage/min": -0.9015698432922363,
"advantage/max": 8.246562004089355,
"time/assemble_training_data": 6.113327264785767,
"time/kl_vs_base": 100.20314812660217,
"kl_policy_base": 0.0006044299807399511,
"time/train": 636.6128726005554,
"time/save_checkpoint": 9.186229705810547,
"time/total": 6474.802463054657
}
[2026-07-09T06:42:12+00:00] job=1812634 node=node-30 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:26:33+00:00] job=1812736 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:27:32+00:00] job=1813133 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T16:26:04+00:00] job=1813134 node=node-4 ngpu=6 ntrain=2 replicas=4 flash_attn=yes