Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step23
ttt-discover-v3-qwen3-8b-ac1-step23 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 22 (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 22 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 22,
"progress/batch": 22,
"optim/lr": 4e-05,
"progress/done_frac": 0.46,
"puct/buffer_size": 360,
"puct/sampled_size": 8,
"puct/T": 11264,
"puct/scale_last": 0.4936516100901467,
"puct/buffer_value/mean": -1.526897008761525,
"puct/buffer_value/std": 0.08706835610341182,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5063483899098538,
"puct/buffer_timestep/mean": 10.244444444444444,
"puct/buffer_timestep/std": 6.4983948635389295,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 21.0,
"puct/buffer_construction_len/mean": 1120.9194444444445,
"puct/buffer_construction_len/std": 703.4592435953141,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.506360266328062,
"puct/sampled_value/std": 6.5335699209280866e-09,
"puct/sampled_value/min": -1.5063602734714756,
"puct/sampled_value/max": -1.5063602588566294,
"puct/sampled_timestep/mean": 21.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 21.0,
"puct/sampled_timestep/max": 21.0,
"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": 4065.6586396694183,
"env/all/ac_tokens_per_turn": 9011.287109375,
"env/all/ob_tokens_per_turn": 2351.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4613779,
"env/all/total_ob_tokens": 1203712,
"env/all/time/sampling_mean": 313.1008692779578,
"env/all/time/sampling_max": 412.7062075138092,
"env/all/time/env_step_mean": 1658.9042533799075,
"env/all/time/env_step_max": 3657.9539301395416,
"env/all/reward/mean": 0.47185555476391716,
"env/all/reward/max": 0.6638518168790436,
"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": 0.47185555476391716,
"env/all/correctness": 0.728515625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.5728806792119796,
"env/all/raw_score/min": 1.506360256815695,
"env/all/raw_score/max": 9.403159725795938,
"env/all/initial_raw_score": -1.5063602663280617,
"env/all/initial_raw_score/min": -1.5063602734714756,
"env/all/initial_raw_score/max": -1.5063602588566294,
"env/all/msg": "Success; raw_score=1.5063602588566294",
"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 modified hybrid approach combining LP optimization, normalization, and adaptive perturbation.\n The algorithm focuses on maintaining normalization of the sequence sum to sqrt(2n) at all steps.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n n = 1000 # Target sequence length\n\n # Initialize with the best sequence from prior knowledge or random\n best_sequence = []\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 best_sequence = np.random.rand(n).tolist()\n sum_initial = sum(best_sequence)\n if sum_initial > 0:\n best_sequence = [x * np.sqrt(2 * n) / sum_initial 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 # Increased number of iterations for better refinement\n t_blend = 0.1 # Larger blend factor for more aggressive updates\n perturb_rate = 0.08 # Rate to perturb the sequence in fallback\n initial_temperature = 1.0\n temperature = initial_temperature # Temperature for perturbation\n cooling_rate = 0.995 # Faster cooling rate for more exploration\n max_perturb_attempts = 20 # Increased number of perturbation attempts per iteration\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 this iteration\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 # Objective: maximize sum of g_opt\n bounds = [(0, 1000.0) for _ in range(n)]\n\n # Solve the LP\n res = linprog(c, A_ub=A_ub, b_ub=b_ub, bounds=bounds, method='highs')\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 # Blend with current sequence and normalize\n new_f = [ (1 - t_blend) * x + t_blend * y for x, y in zip(curr_sequence, new_g) ]\n sum_new_f = np.sum(new_f)\n if sum_new_f > 0:\n scale_factor = np.sqrt(2 * n) / sum_new_f\n new_f = [x * scale_factor for x in new_f]\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 perturbation if LP fails\n for _ in range(max_perturb_attempts):\n # Apply random perturbation\n indices = np.random.choice(n, size=min(n, 5), replace=False)\n perturbation = np.random.rand(n) * perturb_rate * temperature\n curr_sequence = [max(0.0, x + perturbation[i]) for i, x in enumerate(curr_sequence)]\n curr_sequence = [min(1000.0, x) for x in curr_sequence]\n # Normalize the perturbed sequence\n sum_perturbed = np.sum(curr_sequence)\n if sum_perturbed > 0:\n scale_perturbed = np.sqrt(2 * n) / sum_perturbed\n curr_sequence = [x * scale_perturbed 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\n iteration_count += 1\n # Early stopping if no improvement for a while\n if no_improvement_counter > 30:\n print(\"Early stopping due to no improvement\")\n break\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(x) for x in best_sequence]\n```",
"env/all/time/policy": 313.1008692779578,
"env/all/time/policy/min": 158.60280418395996,
"env/all/time/policy/max": 412.7062075138092,
"env/all/time/env_step": 1658.9042533799075,
"env/all/time/env_step/min": 0.6525619029998779,
"env/all/time/env_step/max": 3657.9539301395416,
"env/all/time/reward_compute": 3.273598849773407e-07,
"env/all/time/reward_compute/min": 2.3096799850463867e-07,
"env/all/time/reward_compute/max": 6.146728992462158e-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": 0.010314809158444405,
"advantage/min": -1.0,
"advantage/max": 1.463024616241455,
"time/assemble_training_data": 9.966129541397095,
"time/kl_vs_base": 92.58642864227295,
"kl_policy_base": 0.0007815762073732913,
"time/train": 541.0501577854156,
"time/save_checkpoint": 37.323867321014404,
"time/total": 4749.057883739471
}
[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