Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step47
ttt-discover-v3-qwen3-8b-erdos-step47 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 46 (0-indexed). Strict upstream eval parity: 1100s har…
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 erdos. Checkpoint saved
after training step 46 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 46,
"progress/batch": 46,
"optim/lr": 4e-05,
"progress/done_frac": 0.94,
"puct/buffer_size": 740,
"puct/sampled_size": 8,
"puct/T": 23552,
"puct/scale_last": 0.11905935396483913,
"puct/buffer_value/mean": -0.3830782217300548,
"puct/buffer_value/std": 0.01466426549205312,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.38094064603516675,
"puct/buffer_timestep/mean": 22.164864864864864,
"puct/buffer_timestep/std": 13.410055045863528,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 45.0,
"puct/buffer_construction_len/mean": 79.41486486486487,
"puct/buffer_construction_len/std": 8.19934378442885,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.3809408064551123,
"puct/sampled_value/std": 6.063747929592838e-08,
"puct/sampled_value/min": -0.3809408304543982,
"puct/sampled_value/max": -0.38094064603516675,
"puct/sampled_timestep/mean": 45.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 45.0,
"puct/sampled_timestep/max": 45.0,
"puct/sampled_construction_len/mean": 80.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 80.0,
"puct/sampled_construction_len/max": 80.0,
"time/sampling": 3417.7676634788513,
"env/all/ac_tokens_per_turn": 9247.57421875,
"env/all/ob_tokens_per_turn": 1260.25,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4734758,
"env/all/total_ob_tokens": 645248,
"env/all/time/sampling_mean": 285.4270102768205,
"env/all/time/sampling_max": 398.4766948223114,
"env/all/time/env_step_mean": 1196.1580876223743,
"env/all/time/env_step_max": 3025.2322614192963,
"env/all/reward/mean": 0.43469404861131156,
"env/all/reward/max": 2.6250807944943646,
"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": 0.43469404861131156,
"env/all/correctness": 0.166015625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3821783883610378,
"env/all/raw_score/min": 0.38094064603516675,
"env/all/raw_score/max": 0.48599163204665646,
"env/all/initial_raw_score": -0.3809408064551123,
"env/all/initial_raw_score/min": -0.3809408304543982,
"env/all/initial_raw_score/max": -0.38094064603516675,
"env/all/msg": "C5 mismatch: reported 0.44468841, computed 0.51300496",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n max_corr = np.max(corr[n_points - 1 : 2 * n_points - 1])\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n\n # Define the number of points\n n_points = 128\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n\n # Generate a random binary initial guess\n h_initial = np.zeros(n_points)\n indices = np.random.permutation(n_points)\n h_initial[indices[:n_points//2]] = 1.0\n\n # Set bounds for each point\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(x):\n h = np.array(x)\n sum_h = np.sum(h)\n # Penalties for constraint violation\n penalty = 1e4 * (sum_h - required_sum) ** 2\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Run differential evolution with refined strategy and parameters\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=30,\n maxiter=1500,\n tol=1e-6,\n mutation=(0.2, 0.8),\n recombination=0.8,\n seed=seed,\n x0=h_initial,\n )\n\n best_h = result_de.x\n best_c5 = evaluate_c5(best_h, dx)\n\n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 285.4270102768205,
"env/all/time/policy/min": 131.58872294425964,
"env/all/time/policy/max": 398.4766948223114,
"env/all/time/env_step": 1196.1580876223743,
"env/all/time/env_step/min": 0.0051457881927490234,
"env/all/time/env_step/max": 3025.2322614192963,
"env/all/time/reward_compute": 2.505257725715637e-07,
"env/all/time/reward_compute/min": 1.7881393432617188e-07,
"env/all/time/reward_compute/max": 3.3155083656311035e-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": 0.038418907672166824,
"advantage/min": -0.9515262842178345,
"advantage/max": 7.753537178039551,
"time/assemble_training_data": 7.785640001296997,
"time/kl_vs_base": 77.70211911201477,
"kl_policy_base": 0.0009919257136061788,
"time/train": 528.2221961021423,
"time/save_checkpoint": 8.000873327255249,
"time/total": 4041.14341878891
}
[2026-07-09T06:24:18+00:00] job=1812624 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T07:31:44+00:00] job=1812955 node=node-1 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T07:59:00+00:00] job=1813123 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:28:32+00:00] job=1813124 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T09:35:59+00:00] job=1813609 node=node-6 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:45:57+00:00] job=1813622 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-10T08:20:19+00:00] job=1813610 node=node-29 ngpu=6 ntrain=2 replicas=4 flash_attn=yes