Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step22
ttt-discover-v3-qwen3-8b-erdos-step22 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 21 (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 21 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 21,
"progress/batch": 21,
"optim/lr": 4e-05,
"progress/done_frac": 0.44,
"puct/buffer_size": 344,
"puct/sampled_size": 8,
"puct/T": 10752,
"puct/scale_last": 0.11905844662528187,
"puct/buffer_value/mean": -0.3850510723848862,
"puct/buffer_value/std": 0.020386073392585653,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.380941553374724,
"puct/buffer_timestep/mean": 9.744186046511627,
"puct/buffer_timestep/std": 6.2098684525475125,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 20.0,
"puct/buffer_construction_len/mean": 78.68313953488372,
"puct/buffer_construction_len/std": 11.935687913030996,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.380941555412417,
"puct/sampled_value/std": 1.046485984401897e-09,
"puct/sampled_value/min": -0.3809415565122119,
"puct/sampled_value/max": -0.380941553374724,
"puct/sampled_timestep/mean": 20.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 20.0,
"puct/sampled_timestep/max": 20.0,
"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": 2885.1040823459625,
"env/all/ac_tokens_per_turn": 9586.1484375,
"env/all/ob_tokens_per_turn": 1298.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4908108,
"env/all/total_ob_tokens": 664576,
"env/all/time/sampling_mean": 508.8189272773452,
"env/all/time/sampling_max": 761.1844050884247,
"env/all/time/env_step_mean": 609.6583209801465,
"env/all/time/env_step_max": 2120.4251761436462,
"env/all/reward/mean": 0.6626947313579119,
"env/all/reward/max": 2.625074548557776,
"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": 0.6626947313579119,
"env/all/correctness": 0.265625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.40414973703931345,
"env/all/raw_score/min": 0.38094155242130456,
"env/all/raw_score/max": 0.5099451885204241,
"env/all/initial_raw_score": -0.38094155541241703,
"env/all/initial_raw_score/min": -0.3809415565122119,
"env/all/initial_raw_score/max": -0.380941553374724,
"env/all/msg": "RuntimeError: Program execution failed: TypeError: evaluate_C5() missing 1 required positional argument: 'dx'",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\n\ndef evaluate_C5(h, dx):\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\ndef generate_random_initial(n_points, required_sum):\n h = np.random.rand(n_points)\n h_sum = np.sum(h)\n return h * (required_sum / h_sum)\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n \n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2 # Required sum for the constraint\n\n def objective_de(h):\n sum_h = np.sum(h)\n penalty = 1e3 * (sum_h - required_sum)**2\n return evaluate_C5(h, dx) + penalty\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n best_result = None\n best_c5 = float('inf')\n\n # Try multiple initial guesses\n initial_guesses = [\n initial_h_values.copy(),\n generate_random_initial(n_points, required_sum)\n ]\n\n for guess in initial_guesses:\n result_de = differential_evolution(\n objective_de,\n bounds,\n strategy='rand1bin',\n popsize=80,\n maxiter=1500,\n tol=1e-6,\n mutation=(0.5, 1.0),\n recombination=0.9,\n disp=False,\n polish=True,\n x0=guess\n )\n\n best_h_de = result_de.x\n\n def constraint_sum(h):\n return np.sum(h) - required_sum\n\n cons = [{'type': 'eq', 'fun': constraint_sum}]\n\n result_local = minimize(\n evaluate_C5,\n best_h_de,\n method='SLSQP',\n bounds=bounds,\n constraints=cons,\n tol=1e-6,\n options={'maxiter': 200, 'disp': False}\n )\n\n best_h = result_local.x\n current_c5 = evaluate_C5(best_h, dx)\n\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_result = best_h\n\n return best_result, best_c5, n_points\n```",
"env/all/time/policy": 508.8189272773452,
"env/all/time/policy/min": 222.142671585083,
"env/all/time/policy/max": 761.1844050884247,
"env/all/time/env_step": 609.6583209801465,
"env/all/time/env_step/min": 0.005640506744384766,
"env/all/time/env_step/max": 2120.4251761436462,
"env/all/time/reward_compute": 3.6926940083503723e-07,
"env/all/time/reward_compute/min": 2.682209014892578e-07,
"env/all/time/reward_compute/max": 4.76837158203125e-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": 0.030129605904221535,
"advantage/min": -0.974246621131897,
"advantage/max": 7.421206474304199,
"time/assemble_training_data": 8.570597171783447,
"time/kl_vs_base": 127.58699202537537,
"kl_policy_base": 0.000823151902295649,
"time/train": 1022.1307535171509,
"time/save_checkpoint": 14.131578922271729,
"time/total": 4058.8897244930267
}
[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