Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step46
ttt-discover-v3-qwen3-8b-erdos-step46 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 45 (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 45 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 45,
"progress/batch": 45,
"optim/lr": 4e-05,
"progress/done_frac": 0.92,
"puct/buffer_size": 724,
"puct/sampled_size": 8,
"puct/T": 23040,
"puct/scale_last": 0.1190592132566401,
"puct/buffer_value/mean": -0.38312545487952393,
"puct/buffer_value/std": 0.014821935781563596,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809407867433658,
"puct/buffer_timestep/mean": 21.66022099447514,
"puct/buffer_timestep/std": 13.115846682746222,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 44.0,
"puct/buffer_construction_len/mean": 79.40193370165746,
"puct/buffer_construction_len/std": 8.288982677618446,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.3809408325341169,
"puct/sampled_value/std": 2.1311440433541052e-09,
"puct/sampled_value/min": -0.3809408359212147,
"puct/sampled_value/max": -0.38094082995162526,
"puct/sampled_timestep/mean": 44.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 44.0,
"puct/sampled_timestep/max": 44.0,
"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": 3573.6659801006317,
"env/all/ac_tokens_per_turn": 9009.212890625,
"env/all/ob_tokens_per_turn": 1252.25,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4612717,
"env/all/total_ob_tokens": 641152,
"env/all/time/sampling_mean": 272.38390039652586,
"env/all/time/sampling_max": 384.7733099460602,
"env/all/time/env_step_mean": 1358.6831515892409,
"env/all/time/env_step_max": 3190.3145995140076,
"env/all/reward/mean": 0.8772848361101108,
"env/all/reward/max": 2.6250807944943646,
"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": 0.8772848361101108,
"env/all/correctness": 0.3359375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3839067497217663,
"env/all/raw_score/min": 0.38094064603516675,
"env/all/raw_score/max": 0.6529409669615491,
"env/all/initial_raw_score": -0.3809408325341169,
"env/all/initial_raw_score/min": -0.3809408359212147,
"env/all/initial_raw_score/max": -0.38094082995162526,
"env/all/msg": "After normalization, h(x) is not in [0, 1]. Range: [0.0, 1.0000001046004883]",
"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 valid_corr = corr[n_points : 2 * n_points - 1]\n max_corr = np.max(valid_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 # Start from initial_h_values if provided\n if 'initial_h_values' in globals() and len(initial_h_values) > 0:\n n_points = len(initial_h_values)\n else:\n n_points = 128 # Increased sample resolution for better accuracy\n dx = 2.0 / n_points\n required_sum = n_points / 2\n initial_guess = np.random.rand(n_points)\n initial_guess = np.clip(initial_guess, 0.0, 1.0)\n initial_guess = (initial_guess / np.sum(initial_guess)) * required_sum\n if 'initial_h_values' in globals() and len(initial_h_values) > 0:\n initial_guess = initial_h_values.copy()\n required_sum = np.sum(initial_guess)\n dx = 2.0 / n_points\n initial_guess = np.clip(initial_guess, 0.0, 1.0)\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(x):\n h = np.array(x)\n max_c5 = evaluate_c5(h, dx)\n # Penalize constraint violation (smaller penalty for better exploration)\n sum_h = np.sum(h)\n penalty = 1e-3 * (sum_h - required_sum) ** 2\n return max_c5 + penalty\n\n # Run differential evolution with improved parameters\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=50,\n maxiter=500, # Reduced for faster convergence\n tol=1e-10,\n mutation=(0.5, 1.0),\n recombination=0.9,\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": 272.38390039652586,
"env/all/time/policy/min": 124.19305658340454,
"env/all/time/policy/max": 384.7733099460602,
"env/all/time/env_step": 1358.6831515892409,
"env/all/time/env_step/min": 0.005579471588134766,
"env/all/time/env_step/max": 3190.3145995140076,
"env/all/time/reward_compute": 3.3853575587272644e-07,
"env/all/time/reward_compute/min": 2.2351741790771484e-07,
"env/all/time/reward_compute/max": 4.470348358154297e-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": 0.026084912940859795,
"advantage/min": -1.0,
"advantage/max": 4.871953010559082,
"time/assemble_training_data": 6.700545072555542,
"time/kl_vs_base": 81.32971620559692,
"kl_policy_base": 0.0009650640422478318,
"time/train": 513.0658543109894,
"time/save_checkpoint": 13.537063121795654,
"time/total": 4189.608247995377
}
[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