Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step49
ttt-discover-v3-qwen3-8b-ac1-step49 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 48 (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 48 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 48,
"progress/batch": 48,
"optim/lr": 4e-05,
"progress/done_frac": 0.98,
"puct/buffer_size": 774,
"puct/sampled_size": 8,
"puct/T": 24576,
"puct/scale_last": 251.1762954598411,
"puct/buffer_value/mean": -1.8593854555274727,
"puct/buffer_value/std": 9.02217869220005,
"puct/buffer_value/min": -252.6826286848707,
"puct/buffer_value/max": -1.5063332250296204,
"puct/buffer_timestep/mean": 23.191214470284237,
"puct/buffer_timestep/std": 13.976526105853136,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 47.0,
"puct/buffer_construction_len/mean": 1056.2416020671835,
"puct/buffer_construction_len/std": 483.53144539182216,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5063332250296217,
"puct/sampled_value/std": 6.707437923744493e-16,
"puct/sampled_value/min": -1.5063332250296222,
"puct/sampled_value/max": -1.5063332250296204,
"puct/sampled_timestep/mean": 47.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 47.0,
"puct/sampled_timestep/max": 47.0,
"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": 2776.9044756889343,
"env/all/ac_tokens_per_turn": 8120.361328125,
"env/all/ob_tokens_per_turn": 2989.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4157625,
"env/all/total_ob_tokens": 1530368,
"env/all/time/sampling_mean": 281.77463299175724,
"env/all/time/sampling_max": 379.55400824546814,
"env/all/time/env_step_mean": 859.0220431974158,
"env/all/time/env_step_max": 2410.8160791397095,
"env/all/reward/mean": 0.6000445291405057,
"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": 0.6000445291405057,
"env/all/correctness": 0.91796875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.5375200926922228,
"env/all/raw_score/min": 1.5063332250296204,
"env/all/raw_score/max": 2.6666719999999993,
"env/all/initial_raw_score": -1.5063332250296215,
"env/all/initial_raw_score/min": -1.5063332250296222,
"env/all/initial_raw_score/max": -1.5063332250296204,
"env/all/msg": "Success; raw_score=1.5063332250296204",
"env/all/parsed_code": "```python\nimport numpy as np\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 a simulated annealing approach with random exploration and acceptance of sub-optimal steps.\n \"\"\"\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, 0.01)\n best_seq = current_seq.copy()\n best_score = evaluate_sequence(best_seq.tolist())\n n = len(current_seq)\n \n # Simulated Annealing Parameters\n initial_temperature = 100.0\n cooling_rate = 0.995\n max_iterations = 100000\n perturbation_size = 0.05 # Adjust this for larger / smaller changes\n min_temp = 1e-3\n\n current_temp = initial_temperature\n for iteration in range(max_iterations):\n if time.time() > deadline:\n break\n # Perturb the sequence\n perturbed_seq = current_seq.copy()\n # Randomly perturb 2-5 elements to avoid getting stuck\n num_perturb = np.random.randint(2, 5)\n for _ in range(num_perturb):\n idx = np.random.randint(n)\n delta = np.random.normal(0, perturbation_size)\n new_val = perturbed_seq[idx] + delta\n new_val = np.clip(new_val, 0, 1000.0)\n perturbed_seq[idx] = new_val\n \n # Evaluate the new sequence\n try:\n new_score = evaluate_sequence(perturbed_seq.tolist())\n except:\n new_score = np.inf # If any error, treat as invalid\n \n # Update best sequence if better\n if new_score < best_score:\n best_score = new_score\n best_seq = perturbed_seq.copy()\n print(f\"New best score: {best_score}\")\n else:\n # Accept with probability depending on temperature\n if np.random.rand() < np.exp((best_score - new_score) / current_temp):\n best_seq = perturbed_seq.copy()\n \n # Cool down\n current_temp *= cooling_rate\n if current_temp < min_temp:\n break\n \n # Ensure final sequence is valid\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": 281.77463299175724,
"env/all/time/policy/min": 143.9427559375763,
"env/all/time/policy/max": 379.55400824546814,
"env/all/time/env_step": 859.0220431974158,
"env/all/time/env_step/min": 0.7357733249664307,
"env/all/time/env_step/max": 2410.8160791397095,
"env/all/time/reward_compute": 5.639158189296722e-07,
"env/all/time/reward_compute/min": 1.8998980522155762e-07,
"env/all/time/reward_compute/max": 1.5124678611755371e-06,
"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": 0.002961728721857071,
"advantage/min": -1.0,
"advantage/max": 0.4651162624359131,
"time/assemble_training_data": 9.621352910995483,
"time/kl_vs_base": 95.03986930847168,
"kl_policy_base": 0.0007151150493882596,
"time/train": 529.1516017913818,
"time/save_checkpoint": 13.060739755630493,
"time/total": 3426.8196556568146
}
[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