Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step11
ttt-discover-v3-qwen3-8b-ac1-step11 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 10 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 10, 2026
Repo size
354 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 99%
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 10 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 10,
"progress/batch": 10,
"optim/lr": 4e-05,
"progress/done_frac": 0.22,
"puct/buffer_size": 168,
"puct/sampled_size": 8,
"puct/T": 5120,
"puct/scale_last": 0.3984844618881791,
"puct/buffer_value/mean": -1.5449038602714995,
"puct/buffer_value/std": 0.11716478268623358,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5071385789058913,
"puct/buffer_timestep/mean": 4.238095238095238,
"puct/buffer_timestep/std": 3.0379310300450633,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 9.0,
"puct/buffer_construction_len/mean": 1259.1130952380952,
"puct/buffer_construction_len/std": 1012.2230922084041,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.507180901736299,
"puct/sampled_value/std": 1.5998782920080158e-05,
"puct/sampled_value/min": -1.5071873835845786,
"puct/sampled_value/max": -1.5071385789058913,
"puct/sampled_timestep/mean": 9.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 9.0,
"puct/sampled_timestep/max": 9.0,
"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": 5414.892427206039,
"env/all/ac_tokens_per_turn": 8767.529296875,
"env/all/ob_tokens_per_turn": 3299.5,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4488975,
"env/all/total_ob_tokens": 1689344,
"env/all/time/sampling_mean": 325.4665736630559,
"env/all/time/sampling_max": 434.8894429206848,
"env/all/time/env_step_mean": 2785.5299710766412,
"env/all/time/env_step_max": 4992.288222551346,
"env/all/reward/mean": 0.5787322881601966,
"env/all/reward/max": 0.6637477523176615,
"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": 0.5787322881601966,
"env/all/correctness": 0.9296875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 3.581437956875449,
"env/all/raw_score/min": 1.5065964289456981,
"env/all/raw_score/max": 638.7370351777388,
"env/all/initial_raw_score": -1.5071809017362994,
"env/all/initial_raw_score/min": -1.5071873835845786,
"env/all/initial_raw_score/max": -1.5071385789058913,
"env/all/msg": "Evaluation timed out after 1100 seconds.",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\nimport random\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 through a hybrid approach combining advanced LP solutions, gradient-based refinement, \n and population-based optimization. This method emphasizes global exploration via \n differential evolution combined with local refinement using gradient descent.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n n = 1000\n\n # Initialize with previous best or new seed-based distribution\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 # Use Gaussian-initialized sequence, normalized and clamped\n random_seq = np.random.normal(0.5, 0.2, n)\n random_seq = np.maximum(0.0, np.clip(random_seq, 0, 1000.0))\n random_seq *= np.sqrt(2 * n) / np.sum(random_seq)\n best_sequence = random_seq.tolist()\n \n best_score = evaluate_sequence(best_sequence)\n\n # Define objective function for optimization\n def objective(seq):\n nonlocal best_score\n seq = np.clip(seq, 0.0, 1000.0)\n seq = np.maximum(seq, 1e-8) # Avoid zero sum\n score = evaluate_sequence(seq.tolist())\n if np.isinf(score):\n return 1e10 # Penalize invalid sequences\n best_score = min(best_score, score)\n return score\n\n # Use global optimization with differential evolution\n bounds = [(0.0, 1000.0) for _ in range(n)]\n result = differential_evolution(objective, bounds, strategy='best1bin', popsize=15, tol=1e-6, mutation=(0.5, 1.0), recombination=0.8, maxiter=100, disp=False)\n\n # Extract the best sequence found\n best_sequence = result.x.tolist()\n best_score = evaluate_sequence(best_sequence)\n\n # Local refinement step using gradient descent\n def local_refinement(seq, steps=100, lr=0.1):\n seq = np.array(seq)\n for _ in range(steps):\n grad = np.zeros_like(seq)\n for i in range(len(seq)):\n seq_p = seq.copy()\n seq_p[i] += 1e-4\n score_p = objective(seq_p.tolist())\n seq_m = seq.copy()\n seq_m[i] -= 1e-4\n score_m = objective(seq_m.tolist())\n grad[i] = (score_p - score_m) / (2e-4)\n seq -= lr * grad\n seq = np.clip(seq, 0.0, 1000.0)\n seq = np.maximum(seq, 1e-8)\n return seq.tolist()\n\n best_sequence = local_refinement(best_sequence)\n\n # Ensure the final sequence is within bounds and valid\n best_sequence = [max(0.0, min(1000.0, x)) for x in best_sequence]\n final_score = evaluate_sequence(best_sequence)\n print(f\"Final sequence score: {final_score}\")\n return best_sequence\n```",
"env/all/time/policy": 325.4665736630559,
"env/all/time/policy/min": 144.2449016571045,
"env/all/time/policy/max": 434.8894429206848,
"env/all/time/env_step": 2785.5299710766412,
"env/all/time/env_step/min": 0.008414983749389648,
"env/all/time/env_step/max": 4992.288222551346,
"env/all/time/reward_compute": 3.9348378777503967e-07,
"env/all/time/reward_compute/min": 2.7194619178771973e-07,
"env/all/time/reward_compute/max": 7.972121238708496e-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": 0.009597867727279663,
"advantage/min": -1.0,
"advantage/max": 1.4610655307769775,
"time/assemble_training_data": 11.343345642089844,
"time/kl_vs_base": 92.71050977706909,
"kl_policy_base": 0.0007525138207711279,
"time/train": 578.4648604393005,
"time/save_checkpoint": 34.74908232688904,
"time/total": 6135.997016906738
}
[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