Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step30
ttt-discover-v3-qwen3-8b-erdos-step30 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 29 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 16, 2026
Repo size
354 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 99%
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 29 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 29,
"progress/batch": 29,
"optim/lr": 4e-05,
"progress/done_frac": 0.6,
"puct/buffer_size": 472,
"puct/sampled_size": 8,
"puct/T": 14848,
"puct/scale_last": 0.11905872740134948,
"puct/buffer_value/mean": -0.3839546624642774,
"puct/buffer_value/std": 0.01749777926065008,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809412725986564,
"puct/buffer_timestep/mean": 13.745762711864407,
"puct/buffer_timestep/std": 8.518361464823458,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 28.0,
"puct/buffer_construction_len/mean": 79.04025423728814,
"puct/buffer_construction_len/std": 10.206370058246558,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38094150289051043,
"puct/sampled_value/std": 5.492475597973576e-08,
"puct/sampled_value/min": -0.3809415368492742,
"puct/sampled_value/max": -0.3809413965179647,
"puct/sampled_timestep/mean": 28.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 28.0,
"puct/sampled_timestep/max": 28.0,
"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": 4547.061759233475,
"env/all/ac_tokens_per_turn": 9375.076171875,
"env/all/ob_tokens_per_turn": 1311.75,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4800039,
"env/all/total_ob_tokens": 671616,
"env/all/time/sampling_mean": 287.36461812257767,
"env/all/time/sampling_max": 413.1762354373932,
"env/all/time/env_step_mean": 1889.9374012309127,
"env/all/time/env_step_max": 4131.738488912582,
"env/all/reward/mean": 0.32535685489918414,
"env/all/reward/max": 2.6250776321099005,
"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": 0.32535685489918414,
"env/all/correctness": 0.125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3850630928315852,
"env/all/raw_score/min": 0.38094110494761857,
"env/all/raw_score/max": 0.5324019851930517,
"env/all/initial_raw_score": -0.38094150289051043,
"env/all/initial_raw_score/min": -0.3809415368492742,
"env/all/initial_raw_score/max": -0.3809413965179647,
"env/all/msg": "RuntimeError: Program execution failed: ValueError: Some entries in x0 lay outside the specified bounds",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\nfrom scipy.signal import correlate\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n one_minus_h = 1.0 - h\n # Compute cross-correlation (full mode gives all shifts)\n corr = correlate(h, one_minus_h, mode='full')\n # Extract middle section where k shifts have full overlap\n corr_valid = corr[n_points - 1 : 2 * n_points - 1]\n max_c5 = np.max(corr_valid) * dx\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 # Ensure initial_h_values is defined\n if initial_h_values is None:\n raise ValueError(\"initial_h_values must be provided\")\n\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\n\n # Generate random initial guess\n initial_guess = np.random.rand(n_points)\n initial_guess = initial_guess / np.sum(initial_guess) * required_sum\n\n # Define bounds\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 # Use smaller penalty coefficient for better exploration\n penalty = 1e3 * (sum_h - required_sum) ** 2\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Run differential evolution with better exploration\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='rand1bin',\n popsize=150,\n maxiter=200,\n tol=1e-6,\n mutation=(0.5, 0.9),\n recombination=0.9,\n seed=seed,\n x0=initial_guess\n )\n\n best_h_de = result_de.x\n best_c5_de = evaluate_c5(best_h_de, dx)\n\n # Local refinement with SLSQP\n result_local = minimize(\n objective,\n best_h_de,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum}],\n tol=1e-6\n )\n\n best_h_final = result_local.x\n best_c5_final = evaluate_c5(best_h_final, dx)\n\n return best_h_final, best_c5_final, n_points\n```",
"env/all/time/policy": 287.36461812257767,
"env/all/time/policy/min": 127.52042150497437,
"env/all/time/policy/max": 413.1762354373932,
"env/all/time/env_step": 1889.9374012309127,
"env/all/time/env_step/min": 0.00551295280456543,
"env/all/time/env_step/max": 4131.738488912582,
"env/all/time/reward_compute": 2.882443368434906e-07,
"env/all/time/reward_compute/min": 2.384185791015625e-07,
"env/all/time/reward_compute/max": 3.8370490074157715e-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": 0.0514688640832901,
"advantage/min": -0.8233073353767395,
"advantage/max": 14.112634658813477,
"time/assemble_training_data": 6.944744825363159,
"time/kl_vs_base": 81.78278875350952,
"kl_policy_base": 0.0009057764546014369,
"time/train": 540.3426730632782,
"time/save_checkpoint": 13.07912015914917,
"time/total": 5190.487153291702
}
[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