Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step26
ttt-discover-v3-qwen3-8b-ac1-step26 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 25 (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 25 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 25,
"progress/batch": 25,
"optim/lr": 4e-05,
"progress/done_frac": 0.52,
"puct/buffer_size": 408,
"puct/sampled_size": 8,
"puct/T": 12800,
"puct/scale_last": 0.4936516100901467,
"puct/buffer_value/mean": -1.524529843087196,
"puct/buffer_value/std": 0.08204585447815967,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5063483899098538,
"puct/buffer_timestep/mean": 11.745098039215685,
"puct/buffer_timestep/std": 7.36404390518193,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 24.0,
"puct/buffer_construction_len/mean": 1106.6936274509803,
"puct/buffer_construction_len/std": 661.9323885550654,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5063602562479375,
"puct/sampled_value/std": 9.716735752599835e-11,
"puct/sampled_value/min": -1.506360256410309,
"puct/sampled_value/max": -1.506360256154211,
"puct/sampled_timestep/mean": 24.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 24.0,
"puct/sampled_timestep/max": 24.0,
"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": 4381.758203268051,
"env/all/ac_tokens_per_turn": 8832.126953125,
"env/all/ob_tokens_per_turn": 2925.5,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4522049,
"env/all/total_ob_tokens": 1497856,
"env/all/time/sampling_mean": 318.2400021185167,
"env/all/time/sampling_max": 410.0692687034607,
"env/all/time/env_step_mean": 2040.262780684512,
"env/all/time/env_step_max": 3969.4381828308105,
"env/all/reward/mean": 0.6240225571844599,
"env/all/reward/max": 0.6638518174472764,
"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": 0.6240225571844599,
"env/all/correctness": 0.966796875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.580993038379819,
"env/all/raw_score/min": 1.5063602555263058,
"env/all/raw_score/max": 11.573302055999457,
"env/all/initial_raw_score": -1.5063602562479377,
"env/all/initial_raw_score/min": -1.506360256410309,
"env/all/initial_raw_score/max": -1.506360256154211,
"env/all/msg": "Success; raw_score=1.506360256154211",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nfrom scipy.optimize import minimize\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 gradient descent optimized for convolution maxima.\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 try:\n best_sequence = list(height_sequence_1)\n except:\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 = 1000 # Increased for more exploration\n learning_rate = 0.05\n decay_rate = 0.999\n momentum = 0.9\n momentum_buffer = np.zeros(n)\n temperature = 1.0\n cooling_rate = 0.995\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 max_b = np.max(conv)\n if max_b < 1e-9:\n break\n k = np.argmax(conv)\n S = np.sum(curr_sequence)\n\n # Compute gradient for each element\n gradient = np.zeros(n)\n for j in range(n):\n if 0 <= k - j < n:\n d_contribution = 2 * curr_sequence[k - j]\n else:\n d_contribution = 0\n grad_term = (2 * n * (d_contribution * S - 2 * max_b)) / (S ** 3)\n gradient[j] = grad_term\n\n # Gradient descent step with momentum\n momentum_buffer = momentum * momentum_buffer - learning_rate * gradient\n new_sequence = np.array(curr_sequence) + momentum_buffer\n new_sequence = np.clip(new_sequence, 0.0, 1000.0)\n new_sequence = new_sequence.tolist()\n\n # Evaluate the new sequence\n new_score = evaluate_sequence(new_sequence)\n if new_score < best_score:\n best_score = new_score\n curr_sequence = new_sequence\n no_improvement_counter = 0\n print(f\"New best score: {best_score}\")\n else:\n no_improvement_counter += 1\n\n # Adjust learning rate and momentum\n if no_improvement_counter >= 50:\n learning_rate *= decay_rate\n momentum *= 0.99\n no_improvement_counter = 0\n\n iteration_count += 1\n if no_improvement_counter >= 500:\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 curr_sequence]\n```",
"env/all/time/policy": 318.2400021185167,
"env/all/time/policy/min": 142.7221851348877,
"env/all/time/policy/max": 410.0692687034607,
"env/all/time/env_step": 2040.262780684512,
"env/all/time/env_step/min": 0.009479522705078125,
"env/all/time/env_step/max": 3969.4381828308105,
"env/all/time/reward_compute": 3.3015385270118713e-07,
"env/all/time/reward_compute/min": 2.0116567611694336e-07,
"env/all/time/reward_compute/max": 8.009374141693115e-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": 0.0033311666920781136,
"advantage/min": -1.0,
"advantage/max": 0.3404254913330078,
"time/assemble_training_data": 5.989467620849609,
"time/kl_vs_base": 94.94329714775085,
"kl_policy_base": 0.0007937636110000312,
"time/train": 562.2947928905487,
"time/save_checkpoint": 34.75484752655029,
"time/total": 5081.682507514954
}
[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