Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step22
ttt-discover-v3-qwen3-8b-erdos-s2-step22 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 21 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 10, 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 21 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 21,
"progress/batch": 21,
"optim/lr": 4e-05,
"progress/done_frac": 0.44,
"puct/buffer_size": 344,
"puct/sampled_size": 8,
"puct/T": 10752,
"puct/scale_last": 0.11907468882780714,
"puct/buffer_value/mean": -0.38516461192638435,
"puct/buffer_value/std": 0.020123922181419656,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.38092531117219286,
"puct/buffer_timestep/mean": 9.744186046511627,
"puct/buffer_timestep/std": 6.2098684525475125,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 20.0,
"puct/buffer_construction_len/mean": 81.84883720930233,
"puct/buffer_construction_len/std": 24.785600396601467,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38092599401654803,
"puct/sampled_value/std": 3.938848349326405e-07,
"puct/sampled_value/min": -0.3809263205023443,
"puct/sampled_value/max": -0.38092531117219286,
"puct/sampled_timestep/mean": 20.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 20.0,
"puct/sampled_timestep/max": 20.0,
"puct/sampled_construction_len/mean": 100.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 100.0,
"puct/sampled_construction_len/max": 100.0,
"time/sampling": 2917.32040309906,
"env/all/ac_tokens_per_turn": 9175.64453125,
"env/all/ob_tokens_per_turn": 1813.875,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4697930,
"env/all/total_ob_tokens": 928704,
"env/all/time/sampling_mean": 523.855565531645,
"env/all/time/sampling_max": 752.6321158409119,
"env/all/time/env_step_mean": 636.3206234737299,
"env/all/time/env_step_max": 2152.9879858493805,
"env/all/reward/mean": 1.0339570902454973,
"env/all/reward/max": 2.625186543421969,
"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": 1.0339570902454973,
"env/all/correctness": 0.3984375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.38624612961551646,
"env/all/raw_score/min": 0.3809253008148594,
"env/all/raw_score/max": 0.5,
"env/all/initial_raw_score": -0.3809259940165481,
"env/all/initial_raw_score/min": -0.3809263205023443,
"env/all/initial_raw_score/max": -0.38092531117219286,
"env/all/msg": "RuntimeError: Program execution failed: TypeError: differential_evolution() got an unexpected keyword argument 'fun'",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 100 # Small n_points to stay within time budget\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n def objective(h_vals):\n # Ensure the constraint is satisfied\n current_sum = np.sum(h_vals)\n desired_sum = n_points / 2\n delta = desired_sum - current_sum\n adjusted = h_vals.copy()\n adjusted[-1] += delta\n adjusted = np.clip(adjusted, 0.0, 1.0)\n\n # Compute cross-correlation to evaluate C5\n corr = np.correlate(adjusted, 1.0 - adjusted, mode=\"full\")\n c5 = np.max(corr) * dx # Scale by dx to match actual integral\n return c5\n\n # Use differential evolution to minimize C5\n result = differential_evolution(\n fun=objective,\n bounds=bounds,\n strategy='best1bin',\n maxiter=200,\n popsize=20,\n tol=1e-6,\n mutation=(0.5, 1),\n recombination=0.8,\n seed=seed,\n disp=False\n )\n\n best_h = result.x\n best_c5 = result.fun\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 523.855565531645,
"env/all/time/policy/min": 236.1785855293274,
"env/all/time/policy/max": 752.6321158409119,
"env/all/time/env_step": 636.3206234737299,
"env/all/time/env_step/min": 0.00734257698059082,
"env/all/time/env_step/max": 2152.9879858493805,
"env/all/time/reward_compute": 3.25031578540802e-07,
"env/all/time/reward_compute/min": 1.7136335372924805e-07,
"env/all/time/reward_compute/max": 4.5821070671081543e-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": 0.021876897662878036,
"advantage/min": -1.0,
"advantage/max": 4.404855728149414,
"time/assemble_training_data": 8.1089928150177,
"time/kl_vs_base": 130.5106315612793,
"kl_policy_base": 0.0008643254404887557,
"time/train": 1027.6125183105469,
"time/save_checkpoint": 12.869650602340698,
"time/total": 4097.843707084656
}
[2026-07-09T06:24:18+00:00] job=1812626 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:02:36+00:00] job=1813125 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T09:32:13+00:00] job=1813126 node=node-3 ngpu=6 ntrain=2 replicas=4 flash_attn=yes
[2026-07-09T10:04:31+00:00] job=1813623 node=node-31 ngpu=3 ntrain=1 replicas=2 flash_attn=yes