Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-erdos-step41
ttt-discover-v3-qwen3-8b-erdos-step41 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 40 (0-indexed). Strict upstream eval parity: 1100s har…
Downloads · 30 days
0
Access
Public
Updated Jul 11, 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 40 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 40,
"progress/batch": 40,
"optim/lr": 4e-05,
"progress/done_frac": 0.82,
"puct/buffer_size": 646,
"puct/sampled_size": 8,
"puct/T": 20480,
"puct/scale_last": 0.1190591760561926,
"puct/buffer_value/mean": -0.38330190041999734,
"puct/buffer_value/std": 0.015526312344695791,
"puct/buffer_value/min": -0.5236859987110454,
"puct/buffer_value/max": -0.3809408239438133,
"puct/buffer_timestep/mean": 19.20123839009288,
"puct/buffer_timestep/std": 11.680196717967329,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 39.0,
"puct/buffer_construction_len/mean": 79.29876160990712,
"puct/buffer_construction_len/std": 8.734591830281605,
"puct/buffer_construction_len/min": 42.0,
"puct/buffer_construction_len/max": 143.0,
"puct/sampled_value/mean": -0.3809410498748921,
"puct/sampled_value/std": 2.1768248630749543e-07,
"puct/sampled_value/min": -0.380941405098638,
"puct/sampled_value/max": -0.3809408447995746,
"puct/sampled_timestep/mean": 39.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 39.0,
"puct/sampled_timestep/max": 39.0,
"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": 3811.718368291855,
"env/all/ac_tokens_per_turn": 9207.5078125,
"env/all/ob_tokens_per_turn": 1253.375,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4714244,
"env/all/total_ob_tokens": 641728,
"env/all/time/sampling_mean": 283.9129703110084,
"env/all/time/sampling_max": 377.88004755973816,
"env/all/time/env_step_mean": 1515.2474582139403,
"env/all/time/env_step_max": 3430.14311671257,
"env/all/reward/mean": 0.6770965825257909,
"env/all/reward/max": 2.6250794395546455,
"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": 0.6770965825257909,
"env/all/correctness": 0.259765625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 0.38445600655827356,
"env/all/raw_score/min": 0.38094084265840705,
"env/all/raw_score/max": 0.509945190694832,
"env/all/initial_raw_score": -0.38094104987489213,
"env/all/initial_raw_score/min": -0.380941405098638,
"env/all/initial_raw_score/max": -0.3809408447995746,
"env/all/msg": "RuntimeError: Program execution failed: TypeError: evaluate_c5() missing 1 required positional argument: 'dx'",
"env/all/parsed_code": "```python\nimport numpy as np\nfrom scipy.optimize import minimize\nfrom collections import defaultdict\n\ndef evaluate_c5(h, dx):\n n_points = len(h)\n one_minus_h = 1.0 - h\n corr = np.correlate(h, one_minus_h, mode='full')\n # Focus on the central portion of the correlation\n max_corr = np.max(corr[n_points - 1 : 2 * n_points - 1])\n return max_corr * dx\n\ndef get_initial_solution(initial_h_values, n_points, dx):\n required_sum = n_points / 2\n # We need to ensure that the initial solution satisfies sum(h) * dx = 1\n # So we scale the initial guess appropriately\n scaled_initial = initial_h_values / np.sum(initial_h_values) * required_sum\n return scaled_initial\n\ndef run(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n\n # Use the provided initial_h_values if available\n if 'initial_h_values' in globals() and len(initial_h_values) > 0:\n n_points = len(initial_h_values)\n dx = 2.0 / n_points\n required_sum = n_points / 2\n initial_guess = get_initial_solution(initial_h_values, n_points, dx)\n else:\n n_points = 80\n dx = 2.0 / n_points\n required_sum = n_points / 2\n initial_guess = np.zeros(n_points)\n indices = np.random.choice(n_points, size=int(n_points // 2), replace=False)\n initial_guess[indices] = 1.0\n\n bounds = [(0.0, 1.0) for _ in range(n_points)]\n constraint = {'type': 'eq', 'fun': lambda x: np.sum(x) - required_sum}\n\n # First phase: constrained optimization using SLSQP\n result = minimize(\n evaluate_c5,\n initial_guess,\n method='SLSQP',\n bounds=bounds,\n constraints=constraint,\n options={'maxiter': 200, 'ftol': 1e-10, 'eps': 1e-8}\n )\n\n best_h = result.x\n best_c5 = evaluate_c5(best_h, dx)\n\n # Second phase: fine-tune the best solution with local optimization\n # Using a small step gradient update\n def objective_for_tuning(x):\n return evaluate_c5(x, dx)\n\n # Small gradient step (e.g., 5 iterations with step size 0.01)\n for _ in range(5):\n grad = np.zeros_like(best_h)\n for i in range(n_points):\n h_plus = best_h.copy()\n h_plus[i] += 0.01\n h_minus = best_h.copy()\n h_minus[i] -= 0.01\n grad[i] = (objective_for_tuning(h_plus) - objective_for_tuning(h_minus)) / 0.02\n best_h -= 0.005 * grad\n # Ensure values stay in [0,1]\n best_h = np.clip(best_h, 0, 1)\n best_c5 = evaluate_c5(best_h, dx)\n\n return best_h, best_c5, n_points\n```",
"env/all/time/policy": 283.9129703110084,
"env/all/time/policy/min": 123.01212072372437,
"env/all/time/policy/max": 377.88004755973816,
"env/all/time/env_step": 1515.2474582139403,
"env/all/time/env_step/min": 0.004912614822387695,
"env/all/time/env_step/max": 3430.14311671257,
"env/all/time/reward_compute": 2.989545464515686e-07,
"env/all/time/reward_compute/min": 1.601874828338623e-07,
"env/all/time/reward_compute/max": 6.370246410369873e-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": 0.033846840262413025,
"advantage/min": -1.0,
"advantage/max": 7.151432037353516,
"time/assemble_training_data": 5.733678102493286,
"time/kl_vs_base": 77.90512084960938,
"kl_policy_base": 0.00092809007037431,
"time/train": 523.4847719669342,
"time/save_checkpoint": 27.772902250289917,
"time/total": 4448.862716674805
}
[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