Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step31
ttt-discover-v3-qwen3-8b-erdos-step31 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 30 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 10, 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 30 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 30,
"progress/batch": 30,
"optim/lr": 4e-05,
"progress/done_frac": 0.62,
"puct/buffer_size": 488,
"puct/sampled_size": 8,
"puct/T": 15360,
"puct/scale_last": 0.11905889505238731,
"puct/buffer_value/mean": -0.38385587159235407,
"puct/buffer_value/std": 0.017216903117263024,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.38094110494761857,
"puct/buffer_timestep/mean": 14.245901639344263,
"puct/buffer_timestep/std": 8.806956523308148,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 29.0,
"puct/buffer_construction_len/mean": 79.07172131147541,
"puct/buffer_construction_len/std": 10.039113032519843,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38094147736369255,
"puct/sampled_value/std": 1.4080326241112644e-07,
"puct/sampled_value/min": -0.3809415339319996,
"puct/sampled_value/max": -0.38094110494761857,
"puct/sampled_timestep/mean": 29.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 29.0,
"puct/sampled_timestep/max": 29.0,
"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": 4532.201035499573,
"env/all/ac_tokens_per_turn": 9431.78125,
"env/all/ob_tokens_per_turn": 1294.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4829072,
"env/all/total_ob_tokens": 662848,
"env/all/time/sampling_mean": 297.9890408664942,
"env/all/time/sampling_max": 405.2496213912964,
"env/all/time/env_step_mean": 1973.0363439093344,
"env/all/time/env_step_max": 4148.597749710083,
"env/all/reward/mean": 0.30178755404769675,
"env/all/reward/max": 2.625079568517705,
"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": 0.30178755404769675,
"env/all/correctness": 0.1171875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.392914646711335,
"env/all/raw_score/min": 0.3809408239438133,
"env/all/raw_score/max": 0.8045137686026345,
"env/all/initial_raw_score": -0.38094147736369266,
"env/all/initial_raw_score/min": -0.3809415339319996,
"env/all/initial_raw_score/max": -0.38094110494761857,
"env/all/msg": "RuntimeError: Program execution failed: UnboundLocalError: cannot access local variable 'h' where it is not associated with a value",
"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 corr = correlate(h, one_minus_h, mode='full')\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 # Use provided initial_h_values if present, else generate random initial guess\n if initial_h_values is None:\n n_points = 80\n # Generate a random initial guess with values in [0, 1]\n h = np.random.rand(n_points)\n required_sum = n_points / 2.0\n current_sum = np.sum(h)\n if current_sum == 0:\n # Avoid division by zero\n h = np.zeros(n_points)\n else:\n h = h * (required_sum / current_sum)\n else:\n n_points = len(initial_h_values)\n h_initial = np.array(initial_h_values)\n required_sum = n_points / 2.0\n current_sum = np.sum(h_initial)\n h_initial = h_initial * (required_sum / current_sum)\n\n dx = 2.0 / n_points\n required_sum = n_points / 2.0\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 penalty = 1e2 * (sum_h - required_sum) ** 2 # Reduced penalty coefficient\n max_c5 = evaluate_c5(h, dx)\n return max_c5 + penalty\n\n # Global optimization using Differential Evolution with adjusted parameters\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n popsize=150,\n maxiter=500,\n tol=1e-6,\n mutation=(0.6, 0.9),\n recombination=0.7,\n seed=seed,\n x0=h\n )\n\n best_h_de = result_de.x\n best_c5_de = evaluate_c5(best_h_de, dx)\n\n # Refinement with Sequential Least Squares Programming (SLSQP)\n def objective_slsqp(x):\n return evaluate_c5(x, dx)\n\n constraints = [{'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum}]\n result_slsqp = minimize(\n objective_slsqp,\n best_h_de,\n bounds=bounds,\n constraints=constraints,\n method='SLSQP',\n tol=1e-6\n )\n\n best_h_final = result_slsqp.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": 297.9890408664942,
"env/all/time/policy/min": 136.01270389556885,
"env/all/time/policy/max": 405.2496213912964,
"env/all/time/env_step": 1973.0363439093344,
"env/all/time/env_step/min": 0.005466938018798828,
"env/all/time/env_step/max": 4148.597749710083,
"env/all/time/reward_compute": 2.891756594181061e-07,
"env/all/time/reward_compute/min": 2.2351741790771484e-07,
"env/all/time/reward_compute/max": 3.46451997756958e-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": 0.05565056577324867,
"advantage/min": -0.8025169968605042,
"advantage/max": 27.785715103149414,
"time/assemble_training_data": 7.083297252655029,
"time/kl_vs_base": 79.9372615814209,
"kl_policy_base": 0.0009045898914337158,
"time/train": 540.1894598007202,
"time/save_checkpoint": 13.967377662658691,
"time/total": 5176.916339874268
}
[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