Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step10
ttt-discover-v3-qwen3-8b-ac1-step10 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 ac1. Checkpoint saved after training step 9 (0-indexed). Strict upstream eval parity: 1100s hard k…
Downloads · 30 days
0
Access
Public
Updated Jul 16, 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 ac1. Checkpoint saved
after training step 9 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 9,
"progress/batch": 9,
"optim/lr": 4e-05,
"progress/done_frac": 0.2,
"puct/buffer_size": 152,
"puct/sampled_size": 8,
"puct/T": 4608,
"puct/scale_last": 0.3984363152328896,
"puct/buffer_value/mean": -1.5488736854264988,
"puct/buffer_value/std": 0.12250355295627302,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5071867255611808,
"puct/buffer_timestep/mean": 3.736842105263158,
"puct/buffer_timestep/std": 2.7499685216027645,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 8.0,
"puct/buffer_construction_len/mean": 1286.3881578947369,
"puct/buffer_construction_len/std": 1060.4888026986605,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5071937883315671,
"puct/sampled_value/std": 8.369598787650677e-06,
"puct/sampled_value/min": -1.5072075072431934,
"puct/sampled_value/max": -1.5071867255611808,
"puct/sampled_timestep/mean": 8.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 8.0,
"puct/sampled_timestep/max": 8.0,
"puct/sampled_construction_len/mean": 1000.0,
"puct/sampled_construction_len/std": 0.0,
"puct/sampled_construction_len/min": 1000.0,
"puct/sampled_construction_len/max": 1000.0,
"time/sampling": 6228.306605577469,
"env/all/ac_tokens_per_turn": 8804.962890625,
"env/all/ob_tokens_per_turn": 3456.875,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4508141,
"env/all/total_ob_tokens": 1769920,
"env/all/time/sampling_mean": 331.83303907932714,
"env/all/time/sampling_max": 426.4717059135437,
"env/all/time/env_step_mean": 2835.311117528472,
"env/all/time/env_step_max": 5795.594966650009,
"env/all/reward/mean": 0.5933160289769015,
"env/all/reward/max": 0.6635089880658891,
"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": 0.5933160289769015,
"env/all/correctness": 0.919921875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 7.219598049773186,
"env/all/raw_score/min": 1.5071385789058913,
"env/all/raw_score/max": 2000.0,
"env/all/initial_raw_score": -1.5071937883315667,
"env/all/initial_raw_score/min": -1.5072075072431934,
"env/all/initial_raw_score/max": -1.5071867255611808,
"env/all/msg": "Success; raw_score=1.5071867255611808",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nimport random\nfrom scipy.optimize import minimize\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers that minimizes the evaluation score \n using a combination of genetic algorithm with adaptive selection and directional mutation. \n The algorithm balances exploration and exploitation to find better sequences.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n n = 1000 # Fixed sequence length for optimal coverage\n \n # Initialize best sequence or reuse previous one\n best_sequence = None\n if 'height_sequence_1' in globals() and isinstance(height_sequence_1, (list, np.ndarray)):\n best_sequence = list(height_sequence_1)\n else:\n random_seq = np.random.uniform(0.0, 1.0, n)\n random_seq *= np.sqrt(2 * n) / np.sum(random_seq)\n best_sequence = random_seq.tolist()\n \n best_score = evaluate_sequence(best_sequence)\n \n # Genetic algorithm parameters\n population_size = 30\n generations = 200\n mutation_rate = 0.15\n \n def evaluate(seq):\n return evaluate_sequence(seq)\n \n def generate_initial_population(size):\n population = []\n for _ in range(size):\n seq = np.random.uniform(0.0, 1.0, n)\n seq *= np.sqrt(2 * n) / np.sum(seq)\n population.append(seq.tolist())\n return population\n \n population = generate_initial_population(population_size)\n population.append(best_sequence)\n \n for gen in range(generations):\n if time.time() >= deadline:\n break\n # Evaluate all population members\n scores = [evaluate(seq) if seq is not None else np.inf for seq in population]\n valid_scores = [s for s in scores if s < np.inf]\n if not valid_scores:\n break\n \n # Select parents based on fitness\n top_indices = np.argsort(scores)[:int(population_size * 0.7)]\n parents = [population[i] for i in top_indices]\n \n # Create new population\n new_population = []\n for _ in range(population_size // 2):\n parent1, parent2 = random.sample(parents, 2)\n child = [0.5 * parent1[i] + 0.5 * parent2[i] for i in range(n)]\n \n # Mutation\n for i in range(n):\n if random.random() < mutation_rate:\n delta = np.random.normal(0, 0.05)\n child[i] = max(0.0, child[i] + delta)\n \n # Ensure sum is above threshold\n sum_child = np.sum(child)\n if sum_child < 0.01:\n scale = 0.01 / sum_child\n child = [x * scale for x in child]\n \n new_population.append(child)\n \n # Replace the worst with new population\n combined = population + new_population\n scores_combined = [evaluate(seq) if seq is not None else np.inf for seq in combined]\n \n # Sort by score\n combined_with_scores = sorted(zip(scores_combined, combined), key=lambda x: x[0])\n population = [seq for score, seq in combined_with_scores[:population_size]]\n \n # Update best sequence\n for seq in population:\n try:\n score = evaluate(seq)\n if score < best_score:\n best_score = score\n best_sequence = seq.copy()\n except:\n pass\n \n return [float(max(0.0, x)) for x in best_sequence]\n```",
"env/all/time/policy": 331.83303907932714,
"env/all/time/policy/min": 92.8678195476532,
"env/all/time/policy/max": 426.4717059135437,
"env/all/time/env_step": 2835.311117528472,
"env/all/time/env_step/min": 0.005079507827758789,
"env/all/time/env_step/max": 5795.594966650009,
"env/all/time/reward_compute": 4.721805453300476e-07,
"env/all/time/reward_compute/min": 2.2351741790771484e-07,
"env/all/time/reward_compute/max": 1.6391277313232422e-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": 0.014521763660013676,
"advantage/min": -1.0,
"advantage/max": 21.22342872619629,
"time/assemble_training_data": 6.135725498199463,
"time/kl_vs_base": 95.24805641174316,
"kl_policy_base": 0.0006950795650482178,
"time/train": 590.5464336872101,
"time/save_checkpoint": 35.03823518753052,
"time/total": 6956.999433994293
}
[2026-07-09T06:24:18+00:00] job=1812628 node=node-12 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T08:30:32+00:00] job=1813127 node=node-22 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:07:08+00:00] job=1813128 node=node-6 ngpu=6 ntrain=2 replicas=4 flash_attn=yes