Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step3
ttt-discover-v3-qwen3-8b-erdos-s2-step3 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 2 (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 2 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 2,
"progress/batch": 2,
"optim/lr": 4e-05,
"progress/done_frac": 0.06,
"puct/buffer_size": 40,
"puct/sampled_size": 8,
"puct/T": 1024,
"puct/scale_last": 0.029687358231089755,
"puct/buffer_value/mean": -0.40939677879674113,
"puct/buffer_value/std": 0.04598240174681004,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.38169229552634576,
"puct/buffer_timestep/mean": 0.2,
"puct/buffer_timestep/std": 0.7483314773547883,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 1.0,
"puct/buffer_construction_len/mean": 68.15,
"puct/buffer_construction_len/std": 23.20080817557871,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 100.0,
"puct/sampled_value/mean": -0.3820603369678236,
"puct/sampled_value/std": 0.00018610139307853336,
"puct/sampled_value/min": -0.3822916119006121,
"puct/sampled_value/max": -0.38169229552634576,
"puct/sampled_timestep/mean": 1.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 1.0,
"puct/sampled_timestep/max": 1.0,
"puct/sampled_construction_len/mean": 61.25,
"puct/sampled_construction_len/std": 22.431841208425134,
"puct/sampled_construction_len/min": 40.0,
"puct/sampled_construction_len/max": 99.0,
"time/sampling": 1131.226990222931,
"env/all/ac_tokens_per_turn": 9203.552734375,
"env/all/ob_tokens_per_turn": 1228.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4712219,
"env/all/total_ob_tokens": 629056,
"env/all/time/sampling_mean": 498.0656366823241,
"env/all/time/sampling_max": 714.2100791931152,
"env/all/time/env_step_mean": 14.689579640049487,
"env/all/time/env_step_max": 501.4041893482208,
"env/all/reward/mean": 0.8349528088716376,
"env/all/reward/max": 2.6220929711635588,
"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": 0.8349528088716376,
"env/all/correctness": 0.330078125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3982830319290857,
"env/all/raw_score/min": 0.38137472041478315,
"env/all/raw_score/max": 0.5251746080363763,
"env/all/initial_raw_score": -0.3820603369678236,
"env/all/initial_raw_score/min": -0.3822916119006121,
"env/all/initial_raw_score/max": -0.38169229552634576,
"env/all/msg": "Success; raw_score=0.38169229552634576",
"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 global h_initial\n h_initial = initial_h_values\n n_points = len(h_initial)\n dx = 2.0 / n_points\n\n def objective(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n max_corr = np.max(corr) * dx\n # Constraint penalty term\n penalty = 1e6 * abs(np.sum(h_array) - n_points / 2)\n return max_corr + penalty\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Run differential evolution\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=20, # Reduced population size for efficiency\n maxiter=100, # Limited iterations to fit budget\n tol=1e-8,\n seed=seed\n )\n best_h = result.x\n best_c5 = np.max(np.correlate(best_h, 1 - best_h, mode=\"full\")) * dx\n\n # Try with initial guess\n result_initial = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=20,\n maxiter=100,\n tol=1e-8,\n seed=seed,\n x0=h_initial\n )\n initial_c5 = np.max(np.correlate(result_initial.x, 1 - result_initial.x, mode=\"full\")) * dx\n if initial_c5 < best_c5:\n best_h = result_initial.x\n best_c5 = initial_c5\n\n # Try random initial guesses\n for _ in range(5):\n h_rand = np.random.uniform(0, 1, n_points)\n h_rand[-1] = (n_points / 2) - np.sum(h_rand[:-1])\n h_rand[-1] = np.clip(h_rand[-1], 0, 1)\n result_rand = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=20,\n maxiter=100,\n tol=1e-8,\n seed=seed,\n x0=h_rand\n )\n rand_c5 = np.max(np.correlate(result_rand.x, 1 - result_rand.x, mode=\"full\")) * dx\n if rand_c5 < best_c5:\n best_h = result_rand.x\n best_c5 = rand_c5\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 498.0656366823241,
"env/all/time/policy/min": 249.20011687278748,
"env/all/time/policy/max": 714.2100791931152,
"env/all/time/env_step": 14.689579640049487,
"env/all/time/env_step/min": 0.007807016372680664,
"env/all/time/env_step/max": 501.4041893482208,
"env/all/time/reward_compute": 3.4319236874580383e-07,
"env/all/time/reward_compute/min": 2.3096799850463867e-07,
"env/all/time/reward_compute/max": 4.991888999938965e-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": 0.025243505835533142,
"advantage/min": -1.0,
"advantage/max": 3.1594295501708984,
"time/assemble_training_data": 10.8297700881958,
"time/kl_vs_base": 128.7538046836853,
"kl_policy_base": 0.0005447123548947275,
"time/train": 965.1692969799042,
"time/save_checkpoint": 7.427568435668945,
"time/total": 2245.087999343872
}
[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