Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step38
ttt-discover-v3-qwen3-8b-erdos-step38 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 37 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 11, 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 37 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 37,
"progress/batch": 37,
"optim/lr": 4e-05,
"progress/done_frac": 0.76,
"puct/buffer_size": 599,
"puct/sampled_size": 8,
"puct/T": 18944,
"puct/scale_last": 0.1190591760561926,
"puct/buffer_value/mean": -0.38331581454950325,
"puct/buffer_value/std": 0.015581220729323305,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809408239438133,
"puct/buffer_timestep/mean": 17.726210350584306,
"puct/buffer_timestep/std": 10.824684158951776,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 36.0,
"puct/buffer_construction_len/mean": 79.24373956594324,
"puct/buffer_construction_len/std": 9.068503347331067,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38094134081491715,
"puct/sampled_value/std": 2.4117569259283527e-07,
"puct/sampled_value/min": -0.3809414964274598,
"puct/sampled_value/max": -0.3809408613432135,
"puct/sampled_timestep/mean": 36.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 36.0,
"puct/sampled_timestep/max": 36.0,
"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": 4387.811342954636,
"env/all/ac_tokens_per_turn": 9117.6328125,
"env/all/ob_tokens_per_turn": 1261.75,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4668228,
"env/all/total_ob_tokens": 646016,
"env/all/time/sampling_mean": 275.6218881648965,
"env/all/time/sampling_max": 391.09349060058594,
"env/all/time/env_step_mean": 1895.6882640472613,
"env/all/time/env_step_max": 4008.8444957733154,
"env/all/reward/mean": 0.6202987941211138,
"env/all/reward/max": 2.6250793388842952,
"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": 0.6202987941211138,
"env/all/correctness": 0.23828125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.38556594680350914,
"env/all/raw_score/min": 0.38094085726727944,
"env/all/raw_score/max": 0.7074506871548626,
"env/all/initial_raw_score": -0.3809413408149171,
"env/all/initial_raw_score/min": -0.3809414964274598,
"env/all/initial_raw_score/max": -0.3809408613432135,
"env/all/msg": "Evaluation timed out after 1100 seconds.",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\nfrom math import inf\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n positive_corr = corr[0:n_points]\n negative_corr = corr[n_points:2 * n_points - 1]\n max_corr = np.max(np.concatenate([positive_corr, negative_corr]))\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n # Initialize random seed\n np.random.seed(seed)\n import random\n random.seed(seed)\n \n # Use initial_h_values if available, which is an 80-point pre-defined array\n if 'initial_h_values' in globals():\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n initial_guess = initial_h_values.copy()\n else:\n n_points = 80 # Use 80 points for more precision\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n # Generate a random initial guess with more balance\n initial_guess = np.random.rand(n_points)\n initial_guess /= np.sum(initial_guess) # Normalize\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(x):\n h = np.array(x)\n sum_h = np.sum(h)\n # Penalize deviation from the integral constraint\n penalty = 1e5 * (sum_h - required_sum) ** 2\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Optimized differential evolution strategy\n result = differential_evolution(\n objective,\n bounds,\n strategy='rand1bin',\n popsize=100, # Larger population for better exploration\n maxiter=2000, # More iterations for deeper convergence\n tol=1e-6,\n mutation=(0.8, 1.0), # Larger mutation range\n recombination=0.9, # Higher recombination for diversity\n seed=seed,\n x0=initial_guess\n )\n\n best_h = result.x\n best_c5 = evaluate_c5(best_h, dx)\n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 275.6218881648965,
"env/all/time/policy/min": 119.35547876358032,
"env/all/time/policy/max": 391.09349060058594,
"env/all/time/env_step": 1895.6882640472613,
"env/all/time/env_step/min": 0.005681276321411133,
"env/all/time/env_step/max": 4008.8444957733154,
"env/all/time/reward_compute": 2.8638169169425964e-07,
"env/all/time/reward_compute/min": 2.2351741790771484e-07,
"env/all/time/reward_compute/max": 3.948807716369629e-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": 0.03294410929083824,
"advantage/min": -0.9785569310188293,
"advantage/max": 8.38438892364502,
"time/assemble_training_data": 6.643655776977539,
"time/kl_vs_base": 76.16133737564087,
"kl_policy_base": 0.0009538866579532623,
"time/train": 520.7319936752319,
"time/save_checkpoint": 15.424909114837646,
"time/total": 5008.033239364624
}
[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
[2026-07-10T08:20:19+00:00] job=1813610 node=node-29 ngpu=6 ntrain=2 replicas=4 flash_attn=yes