Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step7
ttt-discover-v3-qwen3-8b-erdos-s2-step7 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 erdos. Checkpoint saved after training step 6 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 9, 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 erdos. Checkpoint saved
after training step 6 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 6,
"progress/batch": 6,
"optim/lr": 4e-05,
"progress/done_frac": 0.14,
"puct/buffer_size": 104,
"puct/sampled_size": 8,
"puct/T": 3072,
"puct/scale_last": 0.03008763726194008,
"puct/buffer_value/mean": -0.3922487436851724,
"puct/buffer_value/std": 0.031575927925362256,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.38129201649549543,
"puct/buffer_timestep/mean": 2.230769230769231,
"puct/buffer_timestep/std": 1.8873606379054302,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 5.0,
"puct/buffer_construction_len/mean": 61.07692307692308,
"puct/buffer_construction_len/std": 21.688658994831442,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 100.0,
"puct/sampled_value/mean": -0.381314065396926,
"puct/sampled_value/std": 1.187627894800896e-05,
"puct/sampled_value/min": -0.38133243359112323,
"puct/sampled_value/max": -0.38129201649549543,
"puct/sampled_timestep/mean": 5.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 5.0,
"puct/sampled_timestep/max": 5.0,
"puct/sampled_construction_len/mean": 58.125,
"puct/sampled_construction_len/std": 13.30824462504353,
"puct/sampled_construction_len/min": 40.0,
"puct/sampled_construction_len/max": 80.0,
"time/sampling": 1789.1224715709686,
"env/all/ac_tokens_per_turn": 9096.26953125,
"env/all/ob_tokens_per_turn": 1484.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4657290,
"env/all/total_ob_tokens": 759808,
"env/all/time/sampling_mean": 496.70768333692104,
"env/all/time/sampling_max": 727.6475183963776,
"env/all/time/env_step_mean": 61.298753215931356,
"env/all/time/env_step_max": 1100.1196262836456,
"env/all/reward/mean": 0.82611128437503,
"env/all/reward/max": 2.6239395371782583,
"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": 0.82611128437503,
"env/all/correctness": 0.32421875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.39470332290176047,
"env/all/raw_score/min": 0.3811063325170931,
"env/all/raw_score/max": 0.5146358621378109,
"env/all/initial_raw_score": -0.3813140653969261,
"env/all/initial_raw_score/min": -0.38133243359112323,
"env/all/initial_raw_score/max": -0.38129201649549543,
"env/all/msg": "C5 mismatch: reported 0.51962483, computed 0.51634279",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\n\ndef run(seed=42, budget_s=1000, **kwargs):\n # Access initial_h_values from global scope\n h_initial = np.copy(initial_h_values)\n n_points = len(h_initial)\n dx = 2.0 / n_points\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(h_vec):\n # Calculate maximum overlap integral\n correlate_result = np.correlate(h_vec, 1.0 - h_vec, mode='full') * dx\n max_overlap = np.max(correlate_result)\n \n # Constraint penalty: sum of h must be n_points / 2\n sum_h = np.sum(h_vec)\n constraint_violation = np.abs(sum_h - n_points / 2)\n \n return max_overlap + 1e6 * constraint_violation # Strong penalty for constraint violation\n\n # Run Differential Evolution for global optimization\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=20,\n mutation=(0.5, 1),\n recombination=0.9,\n tol=1e-10,\n maxiter=500,\n seed=seed\n )\n\n best_h = result.x\n best_c5 = result.fun\n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 496.70768333692104,
"env/all/time/policy/min": 195.85584259033203,
"env/all/time/policy/max": 727.6475183963776,
"env/all/time/env_step": 61.298753215931356,
"env/all/time/env_step/min": 0.005794525146484375,
"env/all/time/env_step/max": 1100.1196262836456,
"env/all/time/reward_compute": 4.614703357219696e-07,
"env/all/time/reward_compute/min": 1.9744038581848145e-07,
"env/all/time/reward_compute/max": 1.1362135410308838e-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": 0.025421032682061195,
"advantage/min": -0.9217279553413391,
"advantage/max": 4.957485675811768,
"time/assemble_training_data": 8.21110463142395,
"time/kl_vs_base": 130.5301115512848,
"kl_policy_base": 0.0006589367403648794,
"time/train": 978.9041996002197,
"time/save_checkpoint": 7.877158164978027,
"time/total": 2915.908133506775
}
[2026-07-09T06:24:18+00:00] job=1812626 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:02:36+00:00] job=1813125 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:32:13+00:00] job=1813126 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T10:04:31+00:00] job=1813623 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=yes