Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step36
ttt-discover-v3-qwen3-8b-erdos-step36 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 35 (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 35 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 35,
"progress/batch": 35,
"optim/lr": 4e-05,
"progress/done_frac": 0.72,
"puct/buffer_size": 567,
"puct/sampled_size": 8,
"puct/T": 17920,
"puct/scale_last": 0.1190591760561926,
"puct/buffer_value/mean": -0.38344981632494424,
"puct/buffer_value/std": 0.01600437041237827,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809408239438133,
"puct/buffer_timestep/mean": 16.72310405643739,
"puct/buffer_timestep/std": 10.243897597295732,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 34.0,
"puct/buffer_construction_len/mean": 79.2010582010582,
"puct/buffer_construction_len/std": 9.319063104830628,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38094150099236884,
"puct/sampled_value/std": 1.3530068943893095e-09,
"puct/sampled_value/min": -0.38094150338268595,
"puct/sampled_value/max": -0.3809414988079617,
"puct/sampled_timestep/mean": 34.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 34.0,
"puct/sampled_timestep/max": 34.0,
"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": 4295.121944189072,
"env/all/ac_tokens_per_turn": 9584.0,
"env/all/ob_tokens_per_turn": 1236.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4907008,
"env/all/total_ob_tokens": 633024,
"env/all/time/sampling_mean": 298.13993938220665,
"env/all/time/sampling_max": 429.9800171852112,
"env/all/time/env_step_mean": 1824.0398890874349,
"env/all/time/env_step_max": 3883.120820760727,
"env/all/reward/mean": 0.6579196966428247,
"env/all/reward/max": 2.6250753463627334,
"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": 0.6579196966428247,
"env/all/correctness": 0.251953125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3832608572424082,
"env/all/raw_score/min": 0.38094143664669744,
"env/all/raw_score/max": 0.46482237656364545,
"env/all/initial_raw_score": -0.3809415009923688,
"env/all/initial_raw_score/min": -0.38094150338268595,
"env/all/initial_raw_score/max": -0.3809414988079617,
"env/all/msg": "C5 mismatch: reported 0.36532970, computed 0.42654201",
"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 max_c5 = np.max(corr[:n_points]) * dx # Only consider valid shifts\n return max_c5\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n import random\n random.seed(seed)\n\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2\n\n initial_guess = initial_h_values.copy()\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 # Compute the maximum overlap integral\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Differential Evolution parameters\n result = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=150,\n maxiter=300,\n tol=1e-9,\n mutation=(0.7, 1.0), # Adjusted mutation range\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": 298.13993938220665,
"env/all/time/policy/min": 152.82327675819397,
"env/all/time/policy/max": 429.9800171852112,
"env/all/time/env_step": 1824.0398890874349,
"env/all/time/env_step/min": 0.005723237991333008,
"env/all/time/env_step/max": 3883.120820760727,
"env/all/time/reward_compute": 2.942979335784912e-07,
"env/all/time/reward_compute/min": 1.862645149230957e-07,
"env/all/time/reward_compute/max": 3.650784492492676e-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": 0.030640890821814537,
"advantage/min": -1.0,
"advantage/max": 4.8903961181640625,
"time/assemble_training_data": 7.485664367675781,
"time/kl_vs_base": 80.06604266166687,
"kl_policy_base": 0.0009006510372273624,
"time/train": 548.23801612854,
"time/save_checkpoint": 19.48815655708313,
"time/total": 4952.149699926376
}
[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