Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step25
ttt-discover-v3-qwen3-8b-ac1-step25 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 24 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 11, 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 24 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 24,
"progress/batch": 24,
"optim/lr": 4e-05,
"progress/done_frac": 0.5,
"puct/buffer_size": 392,
"puct/sampled_size": 8,
"puct/T": 12288,
"puct/scale_last": 0.4936516100901467,
"puct/buffer_value/mean": -1.5252714588579503,
"puct/buffer_value/std": 0.08361969505424792,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5063483899098538,
"puct/buffer_timestep/mean": 11.244897959183673,
"puct/buffer_timestep/std": 7.075484009534579,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 23.0,
"puct/buffer_construction_len/mean": 1111.0484693877552,
"puct/buffer_construction_len/std": 674.9479573268579,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5063602571110348,
"puct/sampled_value/std": 6.623941082568497e-10,
"puct/sampled_value/min": -1.506360258127451,
"puct/sampled_value/max": -1.5063602562895604,
"puct/sampled_timestep/mean": 23.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 23.0,
"puct/sampled_timestep/max": 23.0,
"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": 4513.265199422836,
"env/all/ac_tokens_per_turn": 8723.2578125,
"env/all/ob_tokens_per_turn": 2685.25,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4466308,
"env/all/total_ob_tokens": 1374848,
"env/all/time/sampling_mean": 303.93663391424343,
"env/all/time/sampling_max": 400.88133096694946,
"env/all/time/env_step_mean": 2162.143363914918,
"env/all/time/env_step_max": 4109.019560813904,
"env/all/reward/mean": 0.629835718328815,
"env/all/reward/max": 0.6638518171705591,
"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": 0.629835718328815,
"env/all/correctness": 0.97265625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.5864191918844623,
"env/all/raw_score/min": 1.506360256154211,
"env/all/raw_score/max": 16.913483368891928,
"env/all/initial_raw_score": -1.506360257111035,
"env/all/initial_raw_score/min": -1.506360258127451,
"env/all/initial_raw_score/max": -1.5063602562895604,
"env/all/msg": "Success; raw_score=1.5063602562895604",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy.optimize import linprog\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 a refined hybrid approach combining adaptive perturbation and structured LP optimization.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n n = 1000\n\n # Initialize with the best sequence from prior knowledge 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 # Start with a normalized random sequence\n best_sequence = np.random.rand(n).tolist()\n sum_a = np.sum(best_sequence)\n best_sequence = [x / sum_a if sum_a > 0 else 1.0 / n for x in best_sequence]\n\n curr_sequence = best_sequence.copy()\n best_score = evaluate_sequence(curr_sequence)\n no_improvement_counter = 0\n iteration_count = 0\n max_iterations = 500 # More exploration\n t_blend = 0.05 # Slightly increased for better exploration\n perturb_rate = 0.35 # Higher initial perturbation rate\n temperature = 1.0\n cooling_rate = 0.995 # Slower cooling\n max_perturb_attempts = 5 # More attempts for better exploration\n adapt_blend = True\n perturb_size = n // 5 # Larger subset to perturb\n\n while time.time() < deadline and iteration_count < max_iterations:\n try:\n # Compute current convolution and max_b\n conv = np.convolve(curr_sequence, curr_sequence)\n M = np.max(conv)\n sum_a = np.sum(curr_sequence)\n if sum_a < 0.001:\n # Skip if sum is too small\n continue\n\n # Generate LP constraints\n A_ub = []\n for k in range(2 * n - 1):\n row = [0.0] * n\n for j in range(n):\n i = k - j\n if 0 <= i < n:\n row[j] += curr_sequence[i]\n A_ub.append(row)\n b_ub = [M] * (2 * n - 1)\n c = [-1.0] * n\n bounds = [(0, 1000.0) for _ in range(n)]\n\n # Solve the LP\n try:\n res = linprog(c, A_ub=A_ub, b_ub=b_ub, bounds=bounds, method='highs', options={'tol': 1e-8})\n if res.success:\n g_opt = res.x\n sum_g_opt = np.sum(g_opt)\n if sum_g_opt > 0:\n # Rescale to have sum sqrt(2n)\n scale = np.sqrt(2 * n) / sum_g_opt\n new_g = [x * scale for x in g_opt]\n # Evaluate current and proposed sequences\n curr_score = evaluate_sequence(curr_sequence)\n new_score = evaluate_sequence(new_g)\n if new_score < curr_score:\n t_blend = max(0.1, t_blend * 1.1)\n else:\n t_blend = max(0.01, t_blend * 0.9)\n\n # Blend with current sequence\n new_f = [ (1 - t_blend) * x + t_blend * y for x, y in zip(curr_sequence, new_g) ]\n curr_sequence = new_f\n\n # Evaluate the new sequence\n curr_score = evaluate_sequence(curr_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = curr_sequence[:]\n no_improvement_counter = 0\n print(f\"New best score: {best_score}\")\n else:\n no_improvement_counter += 1\n else:\n # No improvement, skip\n pass\n else:\n # Fallback to diversified perturbation\n for _ in range(max_perturb_attempts):\n # Randomly select a subset of indices to perturb\n perturb_indices = np.random.choice(n, size=perturb_size, replace=False)\n for idx in perturb_indices:\n if np.random.rand() < 0.5:\n # Random decrease\n curr_sequence[idx] = max(0.0, curr_sequence[idx] - np.random.rand() * perturb_rate * temperature)\n else:\n # Random increase\n curr_sequence[idx] = min(1000.0, curr_sequence[idx] + np.random.rand() * perturb_rate * temperature)\n # Cap values\n curr_sequence = [max(0.0, min(1000.0, x)) for x in curr_sequence]\n # Evaluate the perturbed sequence\n curr_score = evaluate_sequence(curr_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = curr_sequence[:]\n no_improvement_counter = 0\n print(f\"New best score: {best_score}\")\n break\n temperature *= cooling_rate\n except Exception as e:\n print(f\"Error during LP solve: {e}. Falling back to perturbation.\")\n for _ in range(max_perturb_attempts):\n perturb_indices = np.random.choice(n, size=perturb_size, replace=False)\n for idx in perturb_indices:\n if np.random.rand() < 0.5:\n curr_sequence[idx] = max(0.0, curr_sequence[idx] - np.random.rand() * perturb_rate * temperature)\n else:\n curr_sequence[idx] = min(1000.0, curr_sequence[idx] + np.random.rand() * perturb_rate * temperature)\n curr_sequence = [max(0.0, min(1000.0, x)) for x in curr_sequence]\n curr_score = evaluate_sequence(curr_sequence)\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = curr_sequence[:]\n no_improvement_counter = 0\n print(f\"New best score: {best_score}\")\n break\n temperature *= cooling_rate\n\n iteration_count += 1\n if no_improvement_counter >= 50:\n break # Prevent infinite loop in case of no improvement\n\n except Exception as e:\n print(f\"Error during optimization: {e}. Skipping update.\")\n continue\n\n # After iterations or time runs out, return the best sequence\n return [float(max(0.0, x)) for x in best_sequence]\n```",
"env/all/time/policy": 303.93663391424343,
"env/all/time/policy/min": 170.65201807022095,
"env/all/time/policy/max": 400.88133096694946,
"env/all/time/env_step": 2162.143363914918,
"env/all/time/env_step/min": 1.1151454448699951,
"env/all/time/env_step/max": 4109.019560813904,
"env/all/time/reward_compute": 4.10713255405426e-07,
"env/all/time/reward_compute/min": 2.1979212760925293e-07,
"env/all/time/reward_compute/max": 8.23289155960083e-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": 0.0030330342706292868,
"advantage/min": -1.0,
"advantage/max": 0.25999999046325684,
"time/assemble_training_data": 5.9101057052612305,
"time/kl_vs_base": 86.24245262145996,
"kl_policy_base": 0.0007938668131828308,
"time/train": 542.240024805069,
"time/save_checkpoint": 35.32065725326538,
"time/total": 5184.900435447693
}
[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