Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step6
ttt-discover-v3-qwen3-8b-erdos-step6 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 5 (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 5 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 5,
"progress/batch": 5,
"optim/lr": 4e-05,
"progress/done_frac": 0.12,
"puct/buffer_size": 88,
"puct/sampled_size": 8,
"puct/T": 2560,
"puct/scale_last": 0.024110967257326532,
"puct/buffer_value/mean": -0.3943919025759065,
"puct/buffer_value/std": 0.035899224189565364,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3810200048662187,
"puct/buffer_timestep/mean": 1.7272727272727273,
"puct/buffer_timestep/std": 1.6006197146962735,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 4.0,
"puct/buffer_construction_len/mean": 74.39772727272727,
"puct/buffer_construction_len/std": 22.87343624696127,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.3810766414757276,
"puct/sampled_value/std": 2.800837854832582e-05,
"puct/sampled_value/min": -0.38112532425014467,
"puct/sampled_value/max": -0.3810200048662187,
"puct/sampled_timestep/mean": 4.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 4.0,
"puct/sampled_timestep/max": 4.0,
"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": 2205.6713304519653,
"env/all/ac_tokens_per_turn": 9350.00390625,
"env/all/ob_tokens_per_turn": 1227.25,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4787202,
"env/all/total_ob_tokens": 628352,
"env/all/time/sampling_mean": 502.2285939306021,
"env/all/time/sampling_max": 756.0584301948547,
"env/all/time/env_step_mean": 271.91895649861544,
"env/all/time/env_step_max": 1504.7683272361755,
"env/all/reward/mean": 1.0860171035248478,
"env/all/reward/max": 2.6250186816169787,
"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": 1.0860171035248478,
"env/all/correctness": 0.44140625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.4125045681052481,
"env/all/raw_score/min": 0.38094965980730683,
"env/all/raw_score/max": 0.8343314801288622,
"env/all/initial_raw_score": -0.3810766414757276,
"env/all/initial_raw_score/min": -0.38112532425014467,
"env/all/initial_raw_score/max": -0.3810200048662187,
"env/all/msg": "Evaluation timed out after 1100 seconds.",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import basinhopping\n\ndef generate_initial_guess(n_points, original=None, noise_scale=0.1):\n \"\"\"\n Generate a perturbed initial guess based on an original guess.\n \"\"\"\n if original is None:\n return np.random.uniform(0, 1, n_points)\n h = original.copy()\n noise = noise_scale * np.random.normal(0, 1, n_points)\n h += noise\n h = np.clip(h, 0, 1)\n sum_h = np.sum(h)\n scaling = (n_points / 2) / sum_h\n return h * scaling\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n\n np.random.seed(seed)\n\n # Diversified initial guesses\n initial_guesses = []\n\n # 1. Original initial_h_values\n initial_guesses.append(initial_h_values.copy())\n\n # 2. Random initial guess\n initial_guesses.append(generate_initial_guess(n_points))\n\n # 3. Perturbations of the initial_h_values\n for _ in range(3):\n p_guess = generate_initial_guess(n_points, original=initial_h_values, noise_scale=0.05)\n initial_guesses.append(p_guess)\n\n # 4. Alternating high-low pattern\n alternating_guess = np.zeros(n_points)\n alternating_guess[::2] = 1.0 # Set every other point to 1.0\n sum_alternating = np.sum(alternating_guess)\n scaling = (n_points / 2) / sum_alternating\n alternating_guess *= scaling\n initial_guesses.append(alternating_guess)\n\n # 5. High in first half, low in second\n half = n_points // 2\n high_low_guess = np.zeros(n_points)\n high_low_guess[:half] = 1.0\n sum_high_low = np.sum(high_low_guess)\n scaling = (n_points / 2) / sum_high_low\n high_low_guess *= scaling\n initial_guesses.append(high_low_guess)\n\n # 6. Symmetric (triangular) pattern\n symmetric_guess = np.zeros(n_points)\n for i in range(n_points):\n x = i * dx\n symmetric_guess[i] = 1.0 if x < 0.5 else 0.0\n sum_symmetric = np.sum(symmetric_guess)\n scaling = (n_points / 2) / sum_symmetric\n symmetric_guess *= scaling\n initial_guesses.append(symmetric_guess)\n\n # 7. Periodic guess\n periodic_guess = np.zeros(n_points)\n period = n_points // 4\n for i in range(n_points):\n periodic_guess[i] = 0.5 + 0.5 * np.sin(2 * np.pi * i / period)\n sum_periodic = np.sum(periodic_guess)\n scaling = (n_points / 2) / sum_periodic\n periodic_guess *= scaling\n initial_guesses.append(periodic_guess)\n\n best_result = None\n best_c5 = float('inf')\n\n # Optimization parameters\n for h_initial in initial_guesses:\n h_initial = h_initial.astype(float)\n\n def objective(h):\n h = np.array(h)\n h1 = 1.0 - h\n corr = np.correlate(h, h1, mode='full')\n max_corr = np.max(corr)\n return max_corr * dx\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n cons = {'type': 'eq', 'fun': lambda h: np.sum(h) - n_points / 2}\n\n result = basinhopping(\n objective,\n x0=h_initial,\n niter=1000, # Increase number of iterations\n T=0.05, # Lower temperature for better local search\n stepsize=0.1, # Larger stepsize for exploration\n minimizer_kwargs={\n 'method': 'SLSQP',\n 'bounds': bounds,\n 'constraints': cons,\n 'options': {\n 'ftol': 1e-8, 'maxiter': 500\n }\n }\n )\n\n if result.fun < best_c5:\n best_c5 = result.fun\n best_result = result.x\n\n return best_result, best_c5, n_points\n```",
"env/all/time/policy": 502.2285939306021,
"env/all/time/policy/min": 199.19201588630676,
"env/all/time/policy/max": 756.0584301948547,
"env/all/time/env_step": 271.91895649861544,
"env/all/time/env_step/min": 0.005692958831787109,
"env/all/time/env_step/max": 1504.7683272361755,
"env/all/time/reward_compute": 4.7264620661735535e-07,
"env/all/time/reward_compute/min": 2.4586915969848633e-07,
"env/all/time/reward_compute/max": 1.4901161193847656e-06,
"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": 0.02116333320736885,
"advantage/min": -0.993730366230011,
"advantage/max": 3.0548486709594727,
"time/assemble_training_data": 5.554935693740845,
"time/kl_vs_base": 130.24898314476013,
"kl_policy_base": 0.0006720033707097173,
"time/train": 985.815661907196,
"time/save_checkpoint": 14.284597396850586,
"time/total": 3342.6772949695587
}
[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