Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step5
ttt-discover-v3-qwen3-8b-erdos-s2-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": 0.03001298962826765,
"puct/buffer_value/mean": -0.3970806482696751,
"puct/buffer_value/std": 0.03693623883573264,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.38136666412916786,
"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": 66.25,
"puct/buffer_construction_len/std": 23.07701477902008,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 100.0,
"puct/sampled_value/mean": -0.3814376103962141,
"puct/sampled_value/std": 4.830701756101367e-05,
"puct/sampled_value/min": -0.38150534802494734,
"puct/sampled_value/max": -0.38136666412916786,
"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": 46.5,
"puct/sampled_construction_len/std": 6.5,
"puct/sampled_construction_len/min": 40.0,
"puct/sampled_construction_len/max": 53.0,
"time/sampling": 1587.7665629386902,
"env/all/ac_tokens_per_turn": 9310.12109375,
"env/all/ob_tokens_per_turn": 1354.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4766782,
"env/all/total_ob_tokens": 693568,
"env/all/time/sampling_mean": 506.45940474653617,
"env/all/time/sampling_max": 749.5934002399445,
"env/all/time/env_step_mean": 35.97534461785108,
"env/all/time/env_step_max": 1100.12255692482,
"env/all/reward/mean": 0.8767343003161869,
"env/all/reward/max": 2.6226581239694235,
"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": 0.8767343003161869,
"env/all/correctness": 0.34375,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.39411257511364267,
"env/all/raw_score/min": 0.38129253852572564,
"env/all/raw_score/max": 0.5272679225369892,
"env/all/initial_raw_score": -0.3814376103962141,
"env/all/initial_raw_score/min": -0.38150534802494734,
"env/all/initial_raw_score/max": -0.38136666412916786,
"env/all/msg": "Success; raw_score=0.3821819509472587",
"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 global h_initial\n h_initial = initial_h_values # Provided by external context\n n_points = len(h_initial)\n dx = 2.0 / n_points\n\n def objective_with_penalty(h):\n h_array = np.array(h, dtype=np.float64)\n corr = np.correlate(h_array, 1 - h_array, mode=\"full\")\n max_corr = np.max(corr) * dx\n # Penalty for constraint violation\n penalty = 1e2 * (np.sum(h_array) - n_points / 2) ** 2\n return max_corr + penalty\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n\n # --- Differential Evolution Parameters --- #\n result_de = differential_evolution(\n objective_with_penalty,\n bounds,\n strategy='rand1bin',\n maxiter=200, # Increased for better exploration\n popsize=50, # Larger population size\n tol=1e-5,\n mutation=(0.5, 1), # Mutation factors\n recombination=0.8,\n seed=seed\n )\n\n # --- Generate a second candidate: structured initial guess --- #\n # Structured initial guess: symmetric segments\n structured_guess = np.zeros(n_points)\n for i in range(n_points):\n x = i * dx\n if 0.25 <= x < 0.75:\n structured_guess[i] = 1.0\n elif 0.75 <= x < 1.25:\n structured_guess[i] = 0.5 # Middle section\n else:\n structured_guess[i] = 1.0\n structured_guess[-1] = (n_points / 2) - np.sum(structured_guess[:-1])\n structured_guess[-1] = np.clip(structured_guess[-1], 0.0, 1.0)\n\n # Evaluate both initial guesses\n candidates = [result_de.x, structured_guess]\n\n # --- Refine using SLSQP --- #\n def objective_slsqp(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 constraints = [{'type': 'eq', 'fun': lambda h: np.sum(h) - n_points / 2}]\n bounds_slsqp = [(0.0, 1.0) for _ in range(n_points)]\n\n best_h = None\n best_c5 = float('inf')\n\n # Run SLSQP optimization on both candidates\n for h0 in candidates:\n res = minimize(\n fun=objective_slsqp,\n x0=h0,\n method='SLSQP',\n bounds=bounds_slsqp,\n constraints=constraints,\n options={'ftol': 1e-9, 'maxiter': 300}\n )\n if res.fun < best_c5:\n best_c5 = res.fun\n best_h = res.x\n\n # Perturbation step to escape local optima and refine further\n for _ in range(10):\n h_perturbed = best_h + np.random.uniform(-0.02, 0.02, size=n_points)\n h_perturbed = np.clip(h_perturbed, 0.0, 1.0)\n sum_perturbed = np.sum(h_perturbed[:-1])\n h_perturbed[-1] = max(0.0, min(1.0, n_points / 2 - sum_perturbed))\n res_perturbed = minimize(\n fun=objective_slsqp,\n x0=h_perturbed,\n method='SLSQP',\n bounds=bounds_slsqp,\n constraints=constraints,\n options={'ftol': 1e-9, 'maxiter': 100}\n )\n if res_perturbed.fun < best_c5:\n best_c5 = res_perturbed.fun\n best_h = res_perturbed.x\n\n return (best_h, best_c5, n_points)\n```",
"env/all/time/policy": 506.45940474653617,
"env/all/time/policy/min": 236.5149416923523,
"env/all/time/policy/max": 749.5934002399445,
"env/all/time/env_step": 35.97534461785108,
"env/all/time/env_step/min": 0.005960941314697266,
"env/all/time/env_step/max": 1100.12255692482,
"env/all/time/reward_compute": 4.698522388935089e-07,
"env/all/time/reward_compute/min": 3.166496753692627e-07,
"env/all/time/reward_compute/max": 1.166015863418579e-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": 0.02383837103843689,
"advantage/min": -0.9999906420707703,
"advantage/max": 2.9258384704589844,
"time/assemble_training_data": 8.111438035964966,
"time/kl_vs_base": 132.39560890197754,
"kl_policy_base": 0.0005633496330119669,
"time/train": 991.0089113712311,
"time/save_checkpoint": 8.090380430221558,
"time/total": 2729.18105840683
}
[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