Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step48
ttt-discover-v3-qwen3-8b-ac1-step48 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 47 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 12, 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 47 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 47,
"progress/batch": 47,
"optim/lr": 4e-05,
"progress/done_frac": 0.96,
"puct/buffer_size": 758,
"puct/sampled_size": 8,
"puct/T": 24064,
"puct/scale_last": 251.1762954598411,
"puct/buffer_value/mean": -1.8641881261180877,
"puct/buffer_value/std": 9.116804762099415,
"puct/buffer_value/min": -252.6826286848707,
"puct/buffer_value/max": -1.506333225029621,
"puct/buffer_timestep/mean": 22.688654353562004,
"puct/buffer_timestep/std": 13.683885059676541,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 46.0,
"puct/buffer_construction_len/mean": 1057.428759894459,
"puct/buffer_construction_len/std": 488.5382586339802,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.506333225029622,
"puct/sampled_value/std": 5.661048867003676e-16,
"puct/sampled_value/min": -1.5063332250296229,
"puct/sampled_value/max": -1.506333225029621,
"puct/sampled_timestep/mean": 46.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 46.0,
"puct/sampled_timestep/max": 46.0,
"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": 2829.264080762863,
"env/all/ac_tokens_per_turn": 8080.900390625,
"env/all/ob_tokens_per_turn": 3074.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4137421,
"env/all/total_ob_tokens": 1573888,
"env/all/time/sampling_mean": 280.9525748388842,
"env/all/time/sampling_max": 362.2897412776947,
"env/all/time/env_step_mean": 942.7860862114467,
"env/all/time/env_step_max": 2475.529673576355,
"env/all/reward/mean": 0.5867575192706744,
"env/all/reward/max": 0.663863729980263,
"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": 0.5867575192706744,
"env/all/correctness": 0.90234375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 5.871625966211168,
"env/all/raw_score/min": 1.5063332250296204,
"env/all/raw_score/max": 2000.0,
"env/all/initial_raw_score": -1.5063332250296217,
"env/all/initial_raw_score/min": -1.5063332250296229,
"env/all/initial_raw_score/max": -1.506333225029621,
"env/all/msg": "Success; raw_score=1.506333225029621",
"env/all/parsed_code": "```python\nimport numpy as np\nimport cvxpy as cp\nimport time\n\ndef propose_candidate(seed=42, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score\n using an enhanced iterative linear programming approach with adaptive constraints\n and improved solver efficiency.\n \"\"\"\n # --- Initialization and security checks ---\n np.random.seed(seed)\n deadline = time.time() + 1000 - 10 # Run for up to 1000 seconds\n\n # Start with a safe, non-zero initial sequence\n if 'height_sequence_1' in globals() and isinstance(height_sequence_1, (list, np.ndarray)):\n current_seq = np.array(list(height_sequence_1))\n else:\n n = 1000\n current_seq = np.full(n, 0.1) # Starting with higher initial values for better diversity\n best_seq = current_seq.copy()\n best_score = evaluate_sequence(best_seq.tolist())\n n = len(current_seq)\n\n max_iterations = 500 # Increased for more refinement\n iteration = 0\n\n while iteration < max_iterations and time.time() < deadline:\n iteration += 1\n print(f\"LP Iteration {iteration}\")\n\n # Normalize sequence to target sum\n target_sum = np.sqrt(2 * n) # From the initial norm in the provided method\n sum_current = np.sum(current_seq)\n if sum_current == 0:\n current_seq = np.full(n, 0.01)\n else:\n current_seq = current_seq / sum_current * target_sum\n\n # Compute the current maximum convolution value\n b_sequence = np.convolve(current_seq, current_seq)\n M = np.max(b_sequence)\n\n # Construct the linear programming problem\n b = cp.Variable(n, nonneg=True)\n objective = cp.Maximize(cp.sum(b))\n constraints = []\n\n for x in range(2 * n - 1): # For every position in the convolution\n coeff = np.zeros(n)\n for j in range(n):\n if 0 <= x - j < n:\n coeff[j] = current_seq[x - j]\n constraints.append(cp.sum(coeff * b) <= M)\n\n # Solve the linear program with faster solver\n try:\n prob = cp.Problem(objective, constraints)\n prob.solve(solver=cp.ECOS) # ECOS for speed and robustness\n if prob.status != cp.OPTIMAL:\n # Fallback: reuse current sequence if LP is not solvable optimally\n b_opt = current_seq\n else:\n b_opt = b.value\n # Prevent sum from being too small\n sum_b_opt = np.sum(b_opt)\n if sum_b_opt < 0.01:\n scaling_factor = 0.01 / sum_b_opt\n b_opt = b_opt * scaling_factor\n except:\n # Catch solver errors and fallback to current sequence\n print(\"LP error, using current sequence\")\n b_opt = current_seq\n\n # Rescale to target_sum\n sum_b_opt = np.sum(b_opt)\n rescaled_b = b_opt * (target_sum / sum_b_opt)\n\n # Blend with current sequence using increased stepping factor\n t = 0.1 # Increased blending factor for exploration\n new_seq = (1 - t) * current_seq + t * rescaled_b\n new_seq = np.clip(new_seq, 0, 1000.0) # Ensure bounds\n\n # Evaluate the new sequence\n new_score = evaluate_sequence(new_seq.tolist())\n if new_score < best_score:\n best_score = new_score\n best_seq = new_seq.copy()\n print(f\"New best score: {best_score}\")\n\n # Update for next iteration\n current_seq = new_seq\n\n # Ensure final sequence is valid before returning\n best_seq = np.clip(best_seq, 0, 1000.0)\n return [float(x) for x in best_seq.tolist()]\n```",
"env/all/time/policy": 280.9525748388842,
"env/all/time/policy/min": 151.1264250278473,
"env/all/time/policy/max": 362.2897412776947,
"env/all/time/env_step": 942.7860862114467,
"env/all/time/env_step/min": 0.014589309692382812,
"env/all/time/env_step/max": 2475.529673576355,
"env/all/time/reward_compute": 3.380700945854187e-07,
"env/all/time/reward_compute/min": 2.1979212760925293e-07,
"env/all/time/reward_compute/max": 7.636845111846924e-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": 0.0034752804785966873,
"advantage/min": -1.0,
"advantage/max": 0.3999999761581421,
"time/assemble_training_data": 5.423926115036011,
"time/kl_vs_base": 94.41520762443542,
"kl_policy_base": 0.0006693584145978093,
"time/train": 531.5974612236023,
"time/save_checkpoint": 12.39035964012146,
"time/total": 3475.3899252414703
}
[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
[2026-07-11T13:13:37+00:00] job=1824165 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-12T04:58:24+00:00] job=1827204 node=node-4 ngpu=6 ntrain=2 replicas=4 flash_attn=yes