Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step21
ttt-discover-v3-qwen3-8b-erdos-s2-step21 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 20 (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 20 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 20,
"progress/batch": 20,
"optim/lr": 4e-05,
"progress/done_frac": 0.42,
"puct/buffer_size": 328,
"puct/sampled_size": 8,
"puct/T": 10240,
"puct/scale_last": 0.1190746564752162,
"puct/buffer_value/mean": -0.38537130965831146,
"puct/buffer_value/std": 0.020586607839491688,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.3809253435247838,
"puct/buffer_timestep/mean": 9.24390243902439,
"puct/buffer_timestep/std": 5.921356501659915,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 19.0,
"puct/buffer_construction_len/mean": 80.96341463414635,
"puct/buffer_construction_len/std": 25.048706272006328,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 200.0,
"puct/sampled_value/mean": -0.38092628475494594,
"puct/sampled_value/std": 4.549552028424896e-07,
"puct/sampled_value/min": -0.3809268908887385,
"puct/sampled_value/max": -0.3809253435247838,
"puct/sampled_timestep/mean": 19.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 19.0,
"puct/sampled_timestep/max": 19.0,
"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": 2136.6226513385773,
"env/all/ac_tokens_per_turn": 9274.8671875,
"env/all/ob_tokens_per_turn": 1758.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4748732,
"env/all/total_ob_tokens": 900288,
"env/all/time/sampling_mean": 531.8728075814433,
"env/all/time/sampling_max": 775.0547697544098,
"env/all/time/env_step_mean": 290.81958163762465,
"env/all/time/env_step_max": 1397.3624255657196,
"env/all/reward/mean": 1.0213829450277485,
"env/all/reward/max": 2.6251864720433264,
"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": 1.0213829450277485,
"env/all/correctness": 0.39453125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.3876978896581192,
"env/all/raw_score/min": 0.38092531117219286,
"env/all/raw_score/max": 0.62,
"env/all/initial_raw_score": -0.38092628475494594,
"env/all/initial_raw_score/min": -0.3809268908887385,
"env/all/initial_raw_score/max": -0.3809253435247838,
"env/all/msg": "RuntimeError: Program execution failed: NameError: name 'generate_periodic_binary' is not defined",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import differential_evolution, minimize\n\ndef run(seed=42, budget_s=1000, **kwargs):\n n_points = 200 # Use 200 samples for better resolution (still within time budget)\n dx = 2.0 / n_points\n np.random.seed(seed)\n\n # Helper: Generate random binary with proper sum\n def generate_random_binary(n_points):\n h = np.random.randint(0, 2, size=n_points).astype(float)\n current_sum = np.sum(h)\n target_sum = n_points / 2\n delta = target_sum - current_sum\n for i in range(n_points):\n if delta > 0 and h[i] == 0:\n h[i] = 1.0\n delta -= 1\n elif delta < 0 and h[i] == 1:\n h[i] = 0.0\n delta += 1\n return h\n\n # Generate diverse initial guesses\n initial_guesses = [\n initial_h_values.copy(), # Provided\n generate_random_binary(n_points),\n generate_periodic_binary(n_points, m=4), # Vary m\n generate_periodic_binary(n_points, m=6),\n ]\n\n best_h = None\n best_c5 = float('inf')\n\n for h_initial in initial_guesses:\n h_initial = np.array(h_initial, dtype=np.float64)\n h_initial = np.clip(h_initial, 0.0, 1.0)\n current_sum = np.sum(h_initial)\n desired_sum = n_points / 2\n delta = desired_sum - current_sum\n if delta != 0:\n h_initial[-1] = max(0.0, min(1.0, h_initial[-1] + delta))\n h_initial = np.clip(h_initial, 0.0, 1.0)\n\n # Objective with penalty for constraint\n def objective(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n c5_val = np.max(corr) * dx\n penalty = 1e5 * abs(np.sum(h) - n_points / 2)\n return c5_val + penalty\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # Differential Evolution global optimization\n result_de = differential_evolution(\n objective,\n bounds,\n strategy='best1bin',\n maxiter=100,\n popsize=15,\n tol=1e-6,\n mutation=(0.5, 1),\n recombination=0.8,\n seed=seed\n )\n\n # Refinement stage with SLSQP\n h_refined = result_de.x\n for _ in range(300):\n h_array = np.array(h_refined, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n top_ks = np.argsort(corr)[::-1][:5]\n contributions = np.zeros(n_points)\n for k in top_ks:\n for i in range(n_points):\n if i + k < n_points:\n contributions[i] += h_array[i] * (1 - h_array[i + k])\n else:\n contributions[i] += h_array[i] * 1.0\n sorted_indices = np.argsort(contributions)[::-1]\n top_indices = sorted_indices[:20]\n\n h_perturbed = h_refined.copy()\n for i in top_indices:\n h_perturbed[i] += np.random.uniform(-0.01, 0.01)\n h_perturbed[i] = np.clip(h_perturbed[i], 0.0, 1.0)\n\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - np.sum(h_perturbed[:-1])))\n\n # Refine with SLSQP\n def objective_refine(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n return np.max(corr) * dx\n\n def constraint_func(h):\n return np.sum(h) - n_points / 2\n\n res = minimize(\n fun=objective_refine,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds,\n constraints=[{'type': 'eq', 'fun': constraint_func}],\n options={'ftol': 1e-10, 'maxiter': 200}\n )\n if res.fun < best_c5:\n best_h = res.x\n best_c5 = res.fun\n h_refined = res.x\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 531.8728075814433,
"env/all/time/policy/min": 243.34478402137756,
"env/all/time/policy/max": 775.0547697544098,
"env/all/time/env_step": 290.81958163762465,
"env/all/time/env_step/min": 0.0064237117767333984,
"env/all/time/env_step/max": 1397.3624255657196,
"env/all/time/reward_compute": 4.102475941181183e-07,
"env/all/time/reward_compute/min": 2.4959444999694824e-07,
"env/all/time/reward_compute/max": 1.039355993270874e-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": 0.022852230817079544,
"advantage/min": -1.0,
"advantage/max": 3.1226329803466797,
"time/assemble_training_data": 8.973482608795166,
"time/kl_vs_base": 133.2017364501953,
"kl_policy_base": 0.0008356147445738316,
"time/train": 1036.885672569275,
"time/save_checkpoint": 16.741573095321655,
"time/total": 3336.3140437602997
}
[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