Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step41
ttt-discover-v3-qwen3-8b-erdos-s2-step41 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 40 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 10, 2026
Repo size
355 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 98%
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 40 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 40,
"progress/batch": 40,
"optim/lr": 4e-05,
"progress/done_frac": 0.82,
"puct/buffer_size": 648,
"puct/sampled_size": 8,
"puct/T": 20480,
"puct/scale_last": 0.27910475968829584,
"puct/buffer_value/mean": -0.38419732936684486,
"puct/buffer_value/std": 0.019599369309954737,
"puct/buffer_value/min": -0.66,
"puct/buffer_value/max": -0.3808952403117042,
"puct/buffer_timestep/mean": 19.246913580246915,
"puct/buffer_timestep/std": 11.693124353254769,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 39.0,
"puct/buffer_construction_len/mean": 95.05555555555556,
"puct/buffer_construction_len/std": 31.73049123594014,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.3809171051478373,
"puct/sampled_value/std": 9.66532782769855e-06,
"puct/sampled_value/min": -0.3809236973989164,
"puct/sampled_value/max": -0.3808952403117042,
"puct/sampled_timestep/mean": 39.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 39.0,
"puct/sampled_timestep/max": 39.0,
"puct/sampled_construction_len/mean": 125.0,
"puct/sampled_construction_len/std": 43.30127018922193,
"puct/sampled_construction_len/min": 100.0,
"puct/sampled_construction_len/max": 200.0,
"time/sampling": 1952.6186006069183,
"env/all/ac_tokens_per_turn": 10111.818359375,
"env/all/ob_tokens_per_turn": 1463.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 5177251,
"env/all/total_ob_tokens": 749376,
"env/all/time/sampling_mean": 595.0455849813297,
"env/all/time/sampling_max": 851.5448405742645,
"env/all/time/env_step_mean": 115.30991378473118,
"env/all/time/env_step_max": 1100.123856306076,
"env/all/reward/mean": 0.5864001873998524,
"env/all/reward/max": 2.625394172517621,
"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": 0.5864001873998524,
"env/all/correctness": 0.255859375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.451161320867692,
"env/all/raw_score/min": 0.3808951753660322,
"env/all/raw_score/max": 1.0,
"env/all/initial_raw_score": -0.3809171051478373,
"env/all/initial_raw_score/min": -0.3809236973989164,
"env/all/initial_raw_score/max": -0.3808952403117042,
"env/all/msg": "C5 mismatch: reported 0.00000000, computed 0.99999465",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n n_points = 200 # Sufficient resolution for accuracy\n dx = 2.0 / n_points\n desired_sum = n_points / 2 # Integral of h is 1\n\n # Use initial_h_values if available\n initial_h = np.array(initial_h_values) if 'initial_h_values' in globals() else np.random.rand(n_points)\n\n # If provided initial_h is shorter, interpolate to match n_points\n if len(initial_h) != n_points:\n from scipy.interpolate import interp1d\n interp_func = interp1d(np.linspace(0, 2, len(initial_h)), initial_h, kind='linear', fill_value=\"extrapolate\")\n initial_h = interp_func(np.linspace(0, 2, n_points))\n\n # Adjust initial guess to meet sum constraint and clip to [0, 1]\n current_sum = np.sum(initial_h)\n delta = desired_sum - current_sum\n if abs(delta) > 1e-5:\n initial_h += delta / n_points\n initial_h = np.clip(initial_h, 0.0, 1.0)\n\n best_t = 1.0 # Best known value for t\n best_h = np.zeros(n_points)\n\n def objective(x):\n # Objective is to minimize t\n return x[-1]\n\n def constraint_sum(x):\n # Enforce the integral constraint \u222bh(x)dx = 1\n return np.sum(x[:-1]) - desired_sum\n\n def constraint_overlap(x):\n h = x[:-1]\n t = x[-1]\n overlaps = []\n\n for delta in range(1, n_points):\n len_h = n_points - delta\n if len_h > 0:\n # Compute overlap for positive shift\n overlap = np.sum(h[:len_h] * (1 - h[delta:delta+len_h])) * dx\n overlaps.append(overlap)\n\n return [t - overlap for overlap in overlaps]\n\n cons = [\n {'type': 'eq', 'fun': constraint_sum},\n {'type': 'ineq', 'fun': constraint_overlap}\n ]\n\n bounds = [(0.0, 1.0) for _ in range(n_points)] + [(0.0, 1.0)] # Allow t to be small\n\n try:\n result = minimize(\n fun=objective,\n x0=np.concatenate((initial_h, [0.5])),\n bounds=bounds,\n constraints=cons,\n method='SLSQP',\n options={\n 'ftol': 1e-9,\n 'maxiter': 2000, # More iterations for better convergence\n 'disp': False\n }\n )\n if result.success:\n current_t = result.x[-1]\n if current_t < best_t:\n best_t = current_t\n best_h = result.x[:-1]\n\n except Exception:\n # Handle failure gracefully\n pass\n\n return best_h, best_t, n_points\n```",
"env/all/time/policy": 595.0455849813297,
"env/all/time/policy/min": 281.8510847091675,
"env/all/time/policy/max": 851.5448405742645,
"env/all/time/env_step": 115.30991378473118,
"env/all/time/env_step/min": 0.0050601959228515625,
"env/all/time/env_step/max": 1100.123856306076,
"env/all/time/reward_compute": 4.163011908531189e-07,
"env/all/time/reward_compute/min": 2.7567148208618164e-07,
"env/all/time/reward_compute/max": 8.791685104370117e-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": 0.034546107053756714,
"advantage/min": -0.9733402729034424,
"advantage/max": 7.695284843444824,
"time/assemble_training_data": 6.626867055892944,
"time/kl_vs_base": 135.93713855743408,
"kl_policy_base": 0.0008596000261604786,
"time/train": 1091.6684312820435,
"time/save_checkpoint": 17.765693187713623,
"time/total": 3207.2197346687317
}
[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