Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-s2-step1
ttt-discover-v3-qwen3-8b-erdos-s2-step1 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 0 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 18, 2026
Repo size
356 MB
Likes
0
Public
Click a slice to open those files.
.safetensors349 MB · 98%
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 0 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 0,
"progress/batch": 0,
"optim/lr": 4e-05,
"progress/done_frac": 0.02,
"puct/buffer_size": 8,
"puct/sampled_size": 8,
"puct/T": 0,
"puct/scale_last": 0.02579831795325105,
"puct/buffer_value/mean": -0.5000162594733134,
"puct/buffer_value/std": 0.009304307154703146,
"puct/buffer_value/min": -0.5130522804051018,
"puct/buffer_value/max": -0.4872539624518507,
"puct/buffer_timestep/mean": -1.0,
"puct/buffer_timestep/std": 0.0,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": -1.0,
"puct/buffer_construction_len/mean": 65.75,
"puct/buffer_construction_len/std": 23.13952246698276,
"puct/buffer_construction_len/min": 40.0,
"puct/buffer_construction_len/max": 99.0,
"puct/sampled_value/mean": -0.5000162594733133,
"puct/sampled_value/std": 0.009304307154703146,
"puct/sampled_value/min": -0.5130522804051018,
"puct/sampled_value/max": -0.4872539624518507,
"puct/sampled_timestep/mean": -1.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": -1.0,
"puct/sampled_timestep/max": -1.0,
"puct/sampled_construction_len/mean": 65.75,
"puct/sampled_construction_len/std": 23.13952246698276,
"puct/sampled_construction_len/min": 40.0,
"puct/sampled_construction_len/max": 99.0,
"time/sampling": 1905.2903530597687,
"env/all/ac_tokens_per_turn": 12495.1640625,
"env/all/ob_tokens_per_turn": 613.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 6397524,
"env/all/total_ob_tokens": 313856,
"env/all/time/sampling_mean": 748.7866254849359,
"env/all/time/sampling_max": 1121.0016322135925,
"env/all/time/env_step_mean": 21.169479235075414,
"env/all/time/env_step_max": 1001.5481371879578,
"env/all/reward/mean": 0.5957822815394054,
"env/all/reward/max": 2.6138023705610975,
"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": 0.5957822815394054,
"env/all/correctness": 0.267578125,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.4576090611712783,
"env/all/raw_score/min": 0.38258438553918256,
"env/all/raw_score/max": 0.692576430929055,
"env/all/initial_raw_score": -0.5000162594733133,
"env/all/initial_raw_score/min": -0.5130522804051018,
"env/all/initial_raw_score/max": -0.4872539624518507,
"env/all/msg": "Success; raw_score=0.4730872266189603",
"env/all/parsed_code": "```python\nimport numpy as np\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n \n # Initial h_values provided as a global variable\n h = initial_h_values.copy()\n n_points = len(h)\n dx = 2.0 / n_points\n \n # Function to compute the C\u2085 value\n def compute_c5(h):\n corr = np.correlate(h, 1 - h, mode='full')\n return np.max(corr) * dx\n \n best_h = h.copy()\n best_c5 = compute_c5(best_h)\n history = [best_c5]\n max_iter = 10000 # Adjust based on time budget\n\n for _ in range(max_iter):\n current_c5 = compute_c5(h)\n if current_c5 < best_c5:\n best_c5 = current_c5\n best_h = h.copy()\n history.append(best_c5)\n \n # Compute gradient for the best k\n corr = np.correlate(h, 1 - h, mode='full')\n m = np.argmax(corr)\n s = m - (n_points - 1)\n \n # Compute gradient for shift s\n grad = np.zeros(n_points)\n for j in range(n_points):\n # Term 1: h[j] contributes to (1 - h[j + s])\n j_plus_s = j + s\n if 0 <= j_plus_s < n_points:\n grad[j] += (1 - h[j_plus_s]) * dx\n \n # Term 2: h[j - s] contributes to -h[j]\n j_minus_s = j - s\n if 0 <= j_minus_s < n_points:\n grad[j] += -h[j_minus_s] * dx\n \n # Update h\n learning_rate = 0.01\n h -= learning_rate * grad\n h = np.clip(h, 0.0, 1.0)\n \n # Enforce the integral constraint\n current_sum = np.sum(h)\n if current_sum != 0:\n scale = (n_points / 2) / current_sum\n h = h * scale\n h = np.clip(h, 0.0, 1.0)\n \n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 748.7866254849359,
"env/all/time/policy/min": 362.0082051753998,
"env/all/time/policy/max": 1121.0016322135925,
"env/all/time/env_step": 21.169479235075414,
"env/all/time/env_step/min": 0.007094383239746094,
"env/all/time/env_step/max": 1001.5481371879578,
"env/all/time/reward_compute": 3.0919909477233887e-07,
"env/all/time/reward_compute/min": 1.8998980522155762e-07,
"env/all/time/reward_compute/max": 5.736947059631348e-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": 0.029515601694583893,
"advantage/min": -0.8603240251541138,
"advantage/max": 5.206128120422363,
"time/assemble_training_data": 8.03209400177002,
"time/kl_vs_base": 149.63031196594238,
"kl_policy_base": 0.0004995124181732535,
"time/train": 1269.846137046814,
"time/save_checkpoint": 13.121166706085205,
"time/total": 3351.6880297660828
}
[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