Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step18
ttt-discover-v3-qwen3-8b-erdos-step18 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 17 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 10, 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 17 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 17,
"progress/batch": 17,
"optim/lr": 4e-05,
"progress/done_frac": 0.36,
"puct/buffer_size": 280,
"puct/sampled_size": 8,
"puct/T": 8704,
"puct/scale_last": 0.11905844323878567,
"puct/buffer_value/mean": -0.3859082367383724,
"puct/buffer_value/std": 0.02246720152697283,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809415567612202,
"puct/buffer_timestep/mean": 7.742857142857143,
"puct/buffer_timestep/std": 5.055932058443017,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 16.0,
"puct/buffer_construction_len/mean": 78.38214285714285,
"puct/buffer_construction_len/std": 13.21121151499277,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38094158837328457,
"puct/sampled_value/std": 1.8466627168724218e-08,
"puct/sampled_value/min": -0.3809416035027158,
"puct/sampled_value/max": -0.3809415567612202,
"puct/sampled_timestep/mean": 16.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 16.0,
"puct/sampled_timestep/max": 16.0,
"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": 1598.890146970749,
"env/all/ac_tokens_per_turn": 9060.453125,
"env/all/ob_tokens_per_turn": 1734.5,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4638952,
"env/all/total_ob_tokens": 888064,
"env/all/time/sampling_mean": 500.84921616176143,
"env/all/time/sampling_max": 736.8613419532776,
"env/all/time/env_step_mean": 19.991500078700483,
"env/all/time/env_step_max": 1100.1283185482025,
"env/all/reward/mean": 1.0672474204105886,
"env/all/reward/max": 2.6250745189605085,
"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": 1.0672474204105886,
"env/all/correctness": 0.419921875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3978966775939645,
"env/all/raw_score/min": 0.38094155671635577,
"env/all/raw_score/max": 1.0,
"env/all/initial_raw_score": -0.3809415883732846,
"env/all/initial_raw_score/min": -0.3809416035027158,
"env/all/initial_raw_score/max": -0.3809415567612202,
"env/all/msg": "RuntimeError: Program execution failed: ValueError: Some entries in x0 lay outside the specified bounds",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\nfrom scipy.signal import correlate\n\ndef evaluate_C5(h, dx):\n h1 = 1.0 - h\n corr = correlate(h, h1, mode='full')\n max_corr = np.max(corr)\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n import random\n random.seed(seed)\n\n # Use the provided initial_h_values if available\n n_points = len(initial_h_values) if 'initial_h_values' in globals() else 80\n dx = 2.0 / n_points\n required_sum = int(n_points / 2)\n\n # Define objective function\n def objective(x):\n sum_x = np.sum(x)\n if sum_x == 0:\n return float('inf')\n scaled_x = x * (required_sum / sum_x)\n return evaluate_C5(scaled_x, dx)\n\n # Setup bounds and initial guesses\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Use a random initial guess (not reliant on a fixed one)\n h_random = np.random.rand(n_points)\n h_scaled = h_random * (required_sum / np.sum(h_random))\n initial_guess = h_scaled.copy()\n\n # Run differential_evolution with adjusted parameters\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=20, # Increased population size for diversification\n maxiter=300, # More iterations for refinement\n tol=1e-6,\n mutation=(0.1, 0.5), # Smaller mutation for fine-tuning\n recombination=0.8,\n disp=False,\n polish=True,\n x0=initial_guess\n )\n\n best_h_scaled = result.x * (required_sum / np.sum(result.x))\n best_c5 = result.fun\n return best_h_scaled, best_c5, n_points\n```",
"env/all/time/policy": 500.84921616176143,
"env/all/time/policy/min": 238.74892020225525,
"env/all/time/policy/max": 736.8613419532776,
"env/all/time/env_step": 19.991500078700483,
"env/all/time/env_step/min": 0.005970954895019531,
"env/all/time/env_step/max": 1100.1283185482025,
"env/all/time/reward_compute": 2.915039658546448e-07,
"env/all/time/reward_compute/min": 2.2351741790771484e-07,
"env/all/time/reward_compute/max": 4.6193599700927734e-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": 0.021127987653017044,
"advantage/min": -1.0,
"advantage/max": 2.8725953102111816,
"time/assemble_training_data": 6.064799070358276,
"time/kl_vs_base": 127.62854337692261,
"kl_policy_base": 0.000826895295176655,
"time/train": 1008.4587142467499,
"time/save_checkpoint": 16.456289052963257,
"time/total": 2758.677497625351
}
[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