Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac2-step16
ttt-discover-v3-qwen3-8b-ac2-step16 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 15 (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 15 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 15,
"progress/batch": 15,
"optim/lr": 4e-05,
"progress/done_frac": 0.32,
"puct/buffer_size": 248,
"puct/sampled_size": 8,
"puct/T": 7680,
"puct/scale_last": 0.43720195436973397,
"puct/buffer_value/mean": 0.920194336629851,
"puct/buffer_value/std": 0.057994278307289926,
"puct/buffer_value/min": 0.5041471954736918,
"puct/buffer_value/max": 0.9413491498434258,
"puct/buffer_timestep/mean": 6.741935483870968,
"puct/buffer_timestep/std": 4.479110956828331,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 14.0,
"puct/buffer_construction_len/mean": 2652.0443548387098,
"puct/buffer_construction_len/std": 2718.3998914447525,
"puct/buffer_construction_len/min": 1024.0,
"puct/buffer_construction_len/max": 32768.0,
"puct/sampled_value/mean": 0.9399576797590523,
"puct/sampled_value/std": 0.00026924071716457004,
"puct/sampled_value/min": 0.9396300460000773,
"puct/sampled_value/max": 0.9403743642027313,
"puct/sampled_timestep/mean": 14.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 14.0,
"puct/sampled_timestep/max": 14.0,
"puct/sampled_construction_len/mean": 2048.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 2048.0,
"puct/sampled_construction_len/max": 2048.0,
"time/sampling": 5161.415862083435,
"env/all/ac_tokens_per_turn": 8652.388671875,
"env/all/ob_tokens_per_turn": 4555.25,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4430023,
"env/all/total_ob_tokens": 2332288,
"env/all/time/sampling_mean": 347.7013156991452,
"env/all/time/sampling_max": 459.25940895080566,
"env/all/time/env_step_mean": 2208.2315124813467,
"env/all/time/env_step_max": 4708.883598804474,
"env/all/reward/mean": 0.37974701312307113,
"env/all/reward/max": 0.9413644980551731,
"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": 0.37974701312307113,
"env/all/correctness": 0.431640625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.8797758856063911,
"env/all/raw_score/min": 0.006840612405245569,
"env/all/raw_score/max": 0.9413644980551731,
"env/all/initial_raw_score": 0.9399576797590524,
"env/all/initial_raw_score/min": 0.9396300460000773,
"env/all/initial_raw_score/max": 0.9403743642027313,
"env/all/msg": "Success; raw_score=0.9045037350029277",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom typing import Tuple, List\nimport time\nfrom collections import deque\nimport random\nimport math\n\nnp.seterr(all='ignore')\n\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 if m == 0:\n return 0.0, np.zeros_like(conv)\n dx = 1.0 / (m + 1)\n val = dx * np.sum(np.abs(conv))\n grad = np.sign(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 mask = conv == m\n count = int(mask.sum())\n if count == 0 or m <= 0.0:\n return m, np.zeros_like(conv)\n grad = mask.astype(conv.dtype)\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\nclass _Adam:\n def __init__(self, shape, lr=1e-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 _batch_objective(h_batch: np.ndarray) -> Tuple[np.ndarray, List[np.ndarray]]:\n bsz = h_batch.shape[0]\n c_vals = np.zeros(bsz, dtype=np.float32)\n conv_grads = [None] * bsz\n for b in range(bsz):\n h = np.clip(h_batch[b], 0.0, 1000.0)\n conv = np.convolve(h, h, mode=\"full\")\n c_val, g_conv = _objective_and_grad_conv(conv)\n c_vals[b] = c_val\n conv_grads[b] = g_conv\n return c_vals, conv_grads\n\ndef _phase_update(h_batch, opt_list, lr, add_noise=False, t=0, eta=1e-1, gamma=0.25):\n bsz = h_batch.shape[0]\n c_vals, conv_grads = _batch_objective(h_batch)\n grads = np.zeros_like(h_batch, dtype=h_batch.dtype)\n for b in range(bsz):\n clipped = np.clip(h_batch[b], 0.0, 1000.0)\n grads[b] = _grad_h_from_conv_grad(clipped, conv_grads[b])\n if add_noise:\n sigma = eta / ((t + 1) ** gamma)\n if sigma > 0:\n grads = grads + sigma * np.random.normal(size=grads.shape).astype(grads.dtype)\n for b in range(bsz):\n opt = opt_list[b]\n opt.lr = lr\n h_new = opt.step(h_batch[b], grads[b].astype(h_batch.dtype))\n h_batch[b] = np.clip(h_new, 0.0, 1000.0)\n return h_batch, c_vals\n\ndef _elitist_respawn(h_batch, c_vals, keep_frac, init_sampler, opt_list):\n bsz = h_batch.shape[0]\n keep_n = max(1, int(bsz * keep_frac))\n idx = np.argsort(c_vals)[-keep_n:]\n survivors = h_batch[idx].copy()\n fresh = init_sampler(bsz - keep_n)\n new_batch = np.concatenate([survivors, fresh], axis=0)\n new_opts = [opt_list[i] for i in idx]\n for _ in range(bsz - keep_n):\n new_opts.append(_Adam(shape=h_batch.shape[1:], lr=opt_list[0].lr, dtype=h_batch.dtype))\n return new_batch, new_opts\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=1e-3, steps=500_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, 1000.0)\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, 1000.0)\n last_c = c_val\n return h, float(last_c)\n\ndef construct_function():\n \"\"\"\n Adaptive Pattern Optimization with Evolved Diversity and Adaptive Learning Rates\n Focuses on generating sequences with high-peak structures and refined optimization.\n \"\"\"\n n_start = 128\n n_max = n_start * 4\n bsz = 16\n total_iter = 120_000\n explore_steps = 60_000\n drop_every = 10_000\n keep_frac = 0.3\n lr_explore = 0.1\n lr_exploit = 1e-2\n eta, gamma = 1e-1, 0.25\n\n if 'height_sequence_1' in globals():\n prev = np.array(height_sequence_1, dtype=np.float32)\n else:\n prev = np.ones(n_start, dtype=np.float32)\n\n prev = np.clip(prev, 0.0, 1000.0)\n if prev.shape[0] != n_start:\n x_old = np.linspace(-0.5, 0.5, prev.shape[0])\n x_new = np.linspace(-0.5, 0.5, n_start)\n prev = np.interp(x_new, x_old, prev).astype(np.float32)\n\n def init_sampler(m):\n out = np.random.uniform(0.0, 1000.0, size=(m, n_start)).astype(np.float32)\n for i in range(m):\n if i == 0:\n out[i] = prev\n else:\n if random.random() < 0.3:\n peak_pos = np.random.randint(0, n_start)\n out[i] = np.zeros_like(prev)\n out[i][peak_pos] = 1000.0\n elif random.random() < 0.3:\n peaks = np.random.randint(0, n_start-1, size=5)\n peak_heights = np.random.uniform(0.6, 1000.0, size=5)\n out[i] = np.zeros_like(prev)\n for p, h in zip(peaks, peak_heights):\n if h > 1000.0:\n h = 1000.0\n out[i][p] = h\n elif random.random() < 0.3:\n freq = np.random.uniform(0.1, 0.5)\n phase = np.random.uniform(0, 2 * np.pi)\n t = np.linspace(-0.5, 0.5, n_start)\n out[i] = 0.5 * (1 + np.sin(2 * np.pi * freq * t + phase))\n elif random.random() < 0.1:\n x = np.linspace(-0.5, 0.5, n_start)\n x_center = np.random.uniform(-0.5, 0.5)\n sigma = np.random.uniform(0.05, 0.15)\n out[i] = np.exp(-((x - x_center) ** 2) / (2 * sigma ** 2))\n out[i] = np.minimum(out[i], 1000.0)\n else:\n exp_scale = np.random.uniform(0.25, 0.75)\n out[i] = np.random.exponential(scale=exp_scale, size=n_start)\n return out\n\n h_batch = init_sampler(bsz)\n opt_list = [_Adam(shape=(n_start,), lr=lr_explore, dtype=np.float32) for _ in range(bsz)]\n best_h = h_batch.copy()\n best_c = np.full(bsz, -np.inf, dtype=np.float32)\n\n start_time = time.time()\n\n for t in range(total_iter):\n if t < explore_steps:\n h_batch, c_vals = _phase_update(\n h_batch, opt_list, lr=lr_explore, add_noise=True, t=t, eta=eta, gamma=gamma\n )\n else:\n h_batch, c_vals = _phase_update(\n h_batch, opt_list, lr=lr_exploit, add_noise=False, t=t, eta=eta, gamma=gamma\n )\n\n improved = c_vals > best_c\n best_c = np.where(improved, c_vals, best_c)\n best_h[improved] = h_batch[improved]\n\n if (t + 1) % drop_every == 0:\n h_batch, opt_list = _elitist_respawn(\n h_batch, c_vals, keep_frac=keep_frac, init_sampler=init_sampler, opt_list=opt_list\n )\n\n if t % 1000 == 0:\n elapsed = time.time() - start_time\n print(f\"Iteration {t} (elapsed: {elapsed:.1f}s) - Best score: {best_c[np.argmax(best_c)]:.6f}\")\n if elapsed > 950:\n print(\"Reached time limit, stopping early.\")\n break\n\n idx = np.argmax(best_c)\n h_star = np.clip(best_h[idx].astype(np.float32), 0.0, 1000.0)\n\n h_up1 = _upsample_1d(h_star)\n h_up1, _ = _single_candidate_finetune(h_up1, lr=1e-3, steps=600_000)\n\n h_up2 = _upsample_1d(h_up1)\n h_up2, _ = _single_candidate_finetune(h_up2, lr=1e-3, steps=800_000)\n\n h_up3 = _upsample_1d(h_up2)\n h_up3, _ = _single_candidate_finetune(h_up3, lr=1e-3, steps=1_000_000)\n\n h_final = np.clip(h_up3, 0.0, 1000.0)\n heights = h_final.tolist()\n r_value = evaluate_sequence(heights)\n print(f\"Final C2 lower bound: {r_value:.6f}\")\n return heights\n```",
"env/all/time/policy": 347.7013156991452,
"env/all/time/policy/min": 110.41838216781616,
"env/all/time/policy/max": 459.25940895080566,
"env/all/time/env_step": 2208.2315124813467,
"env/all/time/env_step/min": 0.008156061172485352,
"env/all/time/env_step/max": 4708.883598804474,
"env/all/time/reward_compute": 2.896413207054138e-07,
"env/all/time/reward_compute/min": 1.8998980522155762e-07,
"env/all/time/reward_compute/max": 6.48200511932373e-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": 0.024793699383735657,
"advantage/min": -1.0,
"advantage/max": 5.067156791687012,
"time/assemble_training_data": 10.774739265441895,
"time/kl_vs_base": 109.9005286693573,
"kl_policy_base": 0.0006621094653382897,
"time/train": 648.4151232242584,
"time/save_checkpoint": 29.811079502105713,
"time/total": 5963.195844650269
}
[2026-07-09T06:38:34+00:00] job=1812632 node=node-30 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T06:45:52+00:00] job=1812704 node=node-30 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:00:42+00:00] job=1812735 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:26:33+00:00] job=1812827 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:21:09+00:00] job=1813131 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T14:53:48+00:00] job=1813132 node=node-2 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-10T03:31:51+00:00] job=1816627 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-10T03:54:03+00:00] job=1816628 node=node-29 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-10T07:56:44+00:00] job=1817463 node=node-27 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-10T09:10:26+00:00] job=1817464 node=node-7 ngpu=6 ntrain=2 replicas=4 flash_attn=yes