Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step5
ttt-discover-v3-qwen3-8b-erdos-step5 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 4 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 9, 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 4 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 4,
"progress/batch": 4,
"optim/lr": 4e-05,
"progress/done_frac": 0.1,
"puct/buffer_size": 72,
"puct/sampled_size": 8,
"puct/T": 2048,
"puct/scale_last": 0.02391053111878677,
"puct/buffer_value/mean": -0.3973074881523504,
"puct/buffer_value/std": 0.03909452397446575,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.38122044100475844,
"puct/buffer_timestep/mean": 1.2222222222222223,
"puct/buffer_timestep/std": 1.314684396244359,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 3.0,
"puct/buffer_construction_len/mean": 75.15277777777777,
"puct/buffer_construction_len/std": 24.623644758094596,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.3813806758376017,
"puct/sampled_value/std": 0.00011789979281152868,
"puct/sampled_value/min": -0.38152672264268017,
"puct/sampled_value/max": -0.38122044100475844,
"puct/sampled_timestep/mean": 3.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 3.0,
"puct/sampled_timestep/max": 3.0,
"puct/sampled_construction_len/mean": 71.0,
"puct/sampled_construction_len/std": 11.61895003862225,
"puct/sampled_construction_len/min": 56.0,
"puct/sampled_construction_len/max": 80.0,
"time/sampling": 1791.5235826969147,
"env/all/ac_tokens_per_turn": 9116.861328125,
"env/all/ob_tokens_per_turn": 1379.75,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4667833,
"env/all/total_ob_tokens": 706432,
"env/all/time/sampling_mean": 493.6934704342857,
"env/all/time/sampling_max": 697.0767979621887,
"env/all/time/env_step_mean": 83.86509801028296,
"env/all/time/env_step_max": 1100.1234273910522,
"env/all/reward/mean": 0.9775606049513557,
"env/all/reward/max": 2.6245340427880506,
"env/all/reward/min": 0.0,
"env/all/format": 0.998046875,
"env/all/format/min": 0.0,
"env/all/format/max": 1.0,
"env/all/reward": 0.9775606049513557,
"env/all/correctness": 0.392578125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.4065909937184575,
"env/all/raw_score/min": 0.3810200048662187,
"env/all/raw_score/max": 0.6727673145677004,
"env/all/initial_raw_score": -0.38138067583760166,
"env/all/initial_raw_score/min": -0.38152672264268017,
"env/all/initial_raw_score/max": -0.38122044100475844,
"env/all/msg": "Success; raw_score=0.3812195195929401",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n target_sum = n_points / 2\n\n # Use initial_h_values as the base and scale to meet the target sum\n h_initial = initial_h_values.copy()\n current_sum = np.sum(h_initial)\n scaling_factor = target_sum / current_sum\n h_initial = h_initial * scaling_factor\n h_initial = np.clip(h_initial, 0.0, 1.0)\n\n # Define the objective function\n def objective(h_values):\n h = np.clip(h_values, 0.0, 1.0)\n h1 = 1.0 - h\n corr = np.correlate(h, h1, mode=\"full\")\n return np.max(corr) * dx\n\n # Define bounds for each variable\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Define the equality constraint: sum(h) == target_sum\n def constraint_sum(h_values):\n return np.sum(h_values) - target_sum\n cons = ({'type': 'eq', 'fun': constraint_sum})\n\n # Run optimization with SLSQP\n result = minimize(\n objective,\n x0=h_initial,\n method='SLSQP',\n bounds=bounds,\n constraints=cons,\n options={\n 'ftol': 1e-9,\n 'maxiter': 500,\n 'disp': False\n }\n )\n\n h_values = result.x\n c5_bound = result.fun\n return h_values, c5_bound, n_points\n```",
"env/all/time/policy": 493.6934704342857,
"env/all/time/policy/min": 220.65453791618347,
"env/all/time/policy/max": 697.0767979621887,
"env/all/time/env_step": 83.86509801028296,
"env/all/time/env_step/min": 0.0062220096588134766,
"env/all/time/env_step/max": 1100.1234273910522,
"env/all/time/reward_compute": 7.147900760173798e-07,
"env/all/time/reward_compute/min": 2.123415470123291e-07,
"env/all/time/reward_compute/max": 3.512948751449585e-06,
"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": 0.02399631403386593,
"advantage/min": -1.0,
"advantage/max": 3.3697595596313477,
"time/assemble_training_data": 5.697555065155029,
"time/kl_vs_base": 119.97609543800354,
"kl_policy_base": 0.0006569970864802599,
"time/train": 976.9273540973663,
"time/save_checkpoint": 16.573256731033325,
"time/total": 2911.6897099018097
}
[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