Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step8
ttt-discover-v3-qwen3-8b-erdos-step8 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 7 (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 7 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 7,
"progress/batch": 7,
"optim/lr": 4e-05,
"progress/done_frac": 0.16,
"puct/buffer_size": 120,
"puct/sampled_size": 8,
"puct/T": 3584,
"puct/scale_last": 0.06765834148804023,
"puct/buffer_value/mean": -0.39148749174288683,
"puct/buffer_value/std": 0.031716649705327164,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809460849942446,
"puct/buffer_timestep/mean": 2.7333333333333334,
"puct/buffer_timestep/std": 2.174600857373345,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 6.0,
"puct/buffer_construction_len/mean": 76.05833333333334,
"puct/buffer_construction_len/std": 19.86172862287895,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.38096128332710144,
"puct/sampled_value/std": 9.154134506641186e-06,
"puct/sampled_value/min": -0.3809692161470099,
"puct/sampled_value/max": -0.3809460849942446,
"puct/sampled_timestep/mean": 6.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 6.0,
"puct/sampled_timestep/max": 6.0,
"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": 2187.7080402374268,
"env/all/ac_tokens_per_turn": 9325.58203125,
"env/all/ob_tokens_per_turn": 1343.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4774698,
"env/all/total_ob_tokens": 687936,
"env/all/time/sampling_mean": 504.15072368085384,
"env/all/time/sampling_max": 746.616589307785,
"env/all/time/env_step_mean": 261.5424199150875,
"env/all/time/env_step_max": 1461.189280986786,
"env/all/reward/mean": 0.8064395983320858,
"env/all/reward/max": 2.625065343082106,
"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": 0.8064395983320858,
"env/all/correctness": 0.333984375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.42007229033395627,
"env/all/raw_score/min": 0.3809428882921597,
"env/all/raw_score/max": 0.5250936749749466,
"env/all/initial_raw_score": -0.38096128332710155,
"env/all/initial_raw_score/min": -0.3809692161470099,
"env/all/initial_raw_score/max": -0.3809460849942446,
"env/all/msg": "Evaluation timed out after 1100 seconds.",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import basinhopping, Bounds\nfrom scipy.optimize import minimize\nfrom scipy.signal import correlate\n\ndef generate_initial_guesses(n_points, required_sum):\n guesses = []\n\n # Perturbed guesses from the provided initial_h_values\n for _ in range(8):\n perturbed = initial_h_values.copy() + np.random.normal(0, 0.02, n_points)\n perturbed = np.clip(perturbed, 0.0, 1.0)\n if np.sum(perturbed) != 0:\n scale = required_sum / np.sum(perturbed)\n perturbed *= scale\n guesses.append(perturbed)\n\n # Random guesses with uniform distribution\n for _ in range(10):\n h = np.random.rand(n_points)\n h = h / h.sum() * required_sum\n guesses.append(h)\n\n # Structured initial guess: evenly spaced 1s\n positions = np.linspace(0, 2.0, n_points, endpoint=False)\n selected = np.zeros(n_points)\n selected[::2] = 1\n guesses.append(selected)\n\n # Structured guess: evenly spaced 1s with finer spacing\n even_positions = np.linspace(0, 2.0, 41)\n h_even = np.zeros(n_points)\n dx = 2.0 / n_points\n for i in range(40):\n x = even_positions[i]\n idx = int(x / dx)\n if 0 <= idx < n_points:\n h_even[idx] = 1\n guesses.append(h_even)\n\n return guesses\n\ndef evaluate_C5(h, dx):\n h1 = 1.0 - h\n corr = correlate(h, h1, mode='full')\n max_corr = np.max(corr)\n return max_corr * dx\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = int(n_points / 2)\n\n # Generate diverse initial guesses\n initial_guesses = generate_initial_guesses(n_points, required_sum)\n\n best_h = None\n best_c5 = float('inf')\n\n def objective(h_vec):\n h = np.array(h_vec)\n return evaluate_C5(h, dx)\n\n bounds = Bounds(0.0, 1.0)\n cons = {'type': 'eq', 'fun': lambda h: np.sum(h) - required_sum}\n\n for guess in initial_guesses:\n # Use local optimization for fine tuning\n result = minimize(\n fun=objective,\n x0=guess,\n method='SLSQP',\n bounds=bounds,\n constraints=cons,\n options={'ftol': 1e-8, 'maxiter': 300}\n )\n \n if result.success:\n current_c5 = evaluate_C5(result.x, dx)\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_h = result.x\n print(f\"New best C5: {best_c5:.6f}\")\n\n # Use basinhopping for global search around result\n result_b = basinhopping(\n objective,\n x0=guess,\n niter=200,\n T=0.1,\n stepsize=0.1,\n minimizer_kwargs={\n 'method': 'SLSQP',\n 'bounds': bounds,\n 'constraints': cons,\n 'options': {'ftol': 1e-8, 'maxiter': 200}\n }\n )\n if result_b.fun < best_c5:\n best_c5 = result_b.fun\n best_h = result_b.x\n print(f\"New best C5 (basinhopping): {best_c5:.6f}\")\n\n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 504.15072368085384,
"env/all/time/policy/min": 182.59072279930115,
"env/all/time/policy/max": 746.616589307785,
"env/all/time/env_step": 261.5424199150875,
"env/all/time/env_step/min": 0.006539344787597656,
"env/all/time/env_step/max": 1461.189280986786,
"env/all/time/reward_compute": 3.1385570764541626e-07,
"env/all/time/reward_compute/min": 1.3783574104309082e-07,
"env/all/time/reward_compute/max": 4.991888999938965e-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": 0.027485452592372894,
"advantage/min": -0.9982264041900635,
"advantage/max": 6.599196910858154,
"time/assemble_training_data": 5.8894946575164795,
"time/kl_vs_base": 123.39853763580322,
"kl_policy_base": 0.0007351154345087707,
"time/train": 993.0979363918304,
"time/save_checkpoint": 17.92500591278076,
"time/total": 3329.5551114082336
}
[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