Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step14
ttt-discover-v3-qwen3-8b-erdos-step14 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 13 (0-indexed). Strict upstream eval parity: 1100s har…
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 13 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 13,
"progress/batch": 13,
"optim/lr": 4e-05,
"progress/done_frac": 0.28,
"puct/buffer_size": 216,
"puct/sampled_size": 8,
"puct/T": 6656,
"puct/scale_last": 0.11905831779642695,
"puct/buffer_value/mean": -0.3873695361271381,
"puct/buffer_value/std": 0.02539632481425056,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.38094168220357894,
"puct/buffer_timestep/mean": 5.7407407407407405,
"puct/buffer_timestep/std": 3.9024455067027692,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 12.0,
"puct/buffer_construction_len/mean": 77.90277777777777,
"puct/buffer_construction_len/std": 15.008170357146666,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38094175058023055,
"puct/sampled_value/std": 2.7819369171684798e-08,
"puct/sampled_value/min": -0.3809417690570092,
"puct/sampled_value/max": -0.38094168220357894,
"puct/sampled_timestep/mean": 12.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 12.0,
"puct/sampled_timestep/max": 12.0,
"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": 1353.5142922401428,
"env/all/ac_tokens_per_turn": 9359.212890625,
"env/all/ob_tokens_per_turn": 1560.25,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4791917,
"env/all/total_ob_tokens": 798848,
"env/all/time/sampling_mean": 519.3494763504714,
"env/all/time/sampling_max": 744.257933139801,
"env/all/time/env_step_mean": 18.255104928743094,
"env/all/time/env_step_max": 628.7616131305695,
"env/all/reward/mean": 0.8048719558471631,
"env/all/reward/max": 2.6250741949921754,
"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": 0.8048719558471631,
"env/all/correctness": 0.318359375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3987641513770728,
"env/all/raw_score/min": 0.3809416037295048,
"env/all/raw_score/max": 0.5288281716355826,
"env/all/initial_raw_score": -0.3809417505802305,
"env/all/initial_raw_score/min": -0.3809417690570092,
"env/all/initial_raw_score/max": -0.38094168220357894,
"env/all/msg": "Success; raw_score=0.38208114648015723",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\nfrom scipy.signal import correlate\nimport random\n\ndef evaluate_C5(h, dx):\n h1 = 1.0 - h\n corr = correlate(h, h1, mode='full')\n return np.max(corr) * dx\n\ndef generate_initial_guesses(n_points, required_sum):\n guesses = []\n \n # Binary random initial guess\n for _ in range(20):\n indices = np.random.choice(n_points, required_sum, replace=False)\n h = np.zeros(n_points)\n h[indices] = 1.0\n guesses.append(h)\n \n # Continuous random initial guesses\n for _ in range(20):\n h = np.random.uniform(0, 1, n_points)\n h_scaled = h / np.sum(h) * required_sum\n guesses.append(h_scaled)\n\n return guesses\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n random.seed(seed)\n\n n_points = 48 # Smaller n_points for faster optimization\n dx = 2.0 / n_points\n required_sum = n_points // 2 # Ensure sum(h) == required_sum\n\n guesses = generate_initial_guesses(n_points, required_sum)\n\n best_h = None\n best_c5 = float('inf')\n\n for guess in guesses:\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n cons = [{'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum}]\n\n def objective(x):\n return evaluate_C5(x, dx)\n\n result = minimize(\n fun=objective,\n x0=guess,\n method='SLSQP',\n bounds=bounds,\n constraints=cons,\n options={\n 'ftol': 1e-12,\n 'maxiter': 500, # Sufficient iterations to explore the local region\n 'disp': False\n }\n )\n\n if result.success:\n current_c5 = result.fun\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_h = result.x\n\n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 519.3494763504714,
"env/all/time/policy/min": 251.8583300113678,
"env/all/time/policy/max": 744.257933139801,
"env/all/time/env_step": 18.255104928743094,
"env/all/time/env_step/min": 0.0058956146240234375,
"env/all/time/env_step/max": 628.7616131305695,
"env/all/time/reward_compute": 3.9301812648773193e-07,
"env/all/time/reward_compute/min": 2.3096799850463867e-07,
"env/all/time/reward_compute/max": 9.08970832824707e-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": 0.0264056958258152,
"advantage/min": -1.0,
"advantage/max": 3.8092055320739746,
"time/assemble_training_data": 8.081385612487793,
"time/kl_vs_base": 128.6396918296814,
"kl_policy_base": 0.0008444281993433833,
"time/train": 1020.5820281505585,
"time/save_checkpoint": 11.26515507698059,
"time/total": 2523.7294049263
}
[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