Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-step22
ttt-discover-v3-qwen3-8b-ac1-step22 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 21 (0-indexed). Strict upstream eval parity: 1100s hard…
Downloads · 30 days
0
Access
Public
Updated Jul 10, 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 ac1. Checkpoint saved
after training step 21 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 21,
"progress/batch": 21,
"optim/lr": 4e-05,
"progress/done_frac": 0.44,
"puct/buffer_size": 344,
"puct/sampled_size": 8,
"puct/T": 10752,
"puct/scale_last": 0.39927465088421665,
"puct/buffer_value/mean": -1.5256024902535754,
"puct/buffer_value/std": 0.08407081989669898,
"puct/buffer_value/min": -2.000000000000008,
"puct/buffer_value/max": -1.5063483899098538,
"puct/buffer_timestep/mean": 9.744186046511627,
"puct/buffer_timestep/std": 6.2098684525475125,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 20.0,
"puct/buffer_construction_len/mean": 1126.5436046511627,
"puct/buffer_construction_len/std": 719.1381739380224,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7397.0,
"puct/sampled_value/mean": -1.5063602739485127,
"puct/sampled_value/std": 4.7425049098238336e-09,
"puct/sampled_value/min": -1.5063602760780042,
"puct/sampled_value/max": -1.5063602615266074,
"puct/sampled_timestep/mean": 20.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 20.0,
"puct/sampled_timestep/max": 20.0,
"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": 3103.722046136856,
"env/all/ac_tokens_per_turn": 9435.314453125,
"env/all/ob_tokens_per_turn": 2263.0,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4830881,
"env/all/total_ob_tokens": 1158656,
"env/all/time/sampling_mean": 328.0465613431297,
"env/all/time/sampling_max": 457.1768820285797,
"env/all/time/env_step_mean": 998.7551159900613,
"env/all/time/env_step_max": 2655.849058866501,
"env/all/reward/mean": 0.49522162056972935,
"env/all/reward/max": 0.6638518159796054,
"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": 0.49522162056972935,
"env/all/correctness": 0.806640625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 4.485555464892462,
"env/all/raw_score/min": 1.5063602588566294,
"env/all/raw_score/max": 489.1464679607541,
"env/all/initial_raw_score": -1.5063602739485127,
"env/all/initial_raw_score/min": -1.5063602760780042,
"env/all/initial_raw_score/max": -1.5063602615266074,
"env/all/msg": "Success; raw_score=1.5063602615266074",
"env/all/parsed_code": "```python\nimport numpy as np\nimport random\nimport time\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n \"\"\"\n Proposes a sequence of non-negative numbers to minimize the evaluation score \n using a genetic algorithm approach. It evolves a population of sequences, \n balancing diversity and optimality by applying crossover and mutation, \n gradually converging toward sequences with lower evaluation scores.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n n = 1000\n target_sum = np.sqrt(2 * n)\n \n # Use the provided best_sequence_1 if available\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 best_sequence = np.random.rand(n).tolist()\n sum_a = sum(best_sequence)\n if sum_a > 0:\n best_sequence = [x / sum_a for x in best_sequence]\n best_sequence = [x * target_sum for x in best_sequence]\n \n # Ensure all values are non-negative and within bounds\n best_sequence = [float(max(0.0, min(1000.0, x))) for x in best_sequence]\n best_score = evaluate_sequence(best_sequence)\n \n # Initialize population with diversity, starting from the best sequence\n population_size = 50\n population = []\n for _ in range(population_size):\n if not population:\n seq = best_sequence.copy()\n else:\n seq = np.random.rand(n) \n seq = seq / np.sum(seq) * target_sum\n seq = [float(max(0.0, min(1000.0, x))) for x in seq]\n population.append(seq)\n \n # Evaluate all sequences in the population\n def evaluate(seq):\n return evaluate_sequence(seq)\n \n scores = [evaluate(seq) for seq in population]\n best_idx = np.argmin(scores)\n best_seq = population[best_idx].copy()\n best_score = scores[best_idx]\n \n # Genetic algorithm loop\n generations = 0\n max_generations = 1000\n mutation_rate = 0.1\n crossover_rate = 0.5\n \n while time.time() < deadline and generations < max_generations:\n new_population = []\n for _ in range(population_size // 2):\n # Tournament selection to choose parents\n parent1 = random.choices(population, weights=scores)[0]\n parent2 = random.choices(population, weights=scores)[0]\n \n # Crossover: blend the parents\n child1 = [ (parent1[i] + parent2[i]) / 2 for i in range(n) ]\n child2 = [ (parent1[i] + parent2[i]) / 2 for i in range(n) ]\n \n # Mutation\n for i in range(n):\n if random.random() < mutation_rate:\n mutation = np.random.normal(0, 0.05)\n child1[i] += mutation\n child1[i] = max(0.0, min(1000.0, child1[i]))\n child2[i] += mutation\n child2[i] = max(0.0, min(1000.0, child2[i]))\n \n # Normalize to target_sum\n total = sum(child1)\n if total > 0:\n child1 = [x / total * target_sum for x in child1]\n else:\n child1 = [target_sum / n for _ in range(n)]\n \n total = sum(child2)\n if total > 0:\n child2 = [x / total * target_sum for x in child2]\n else:\n child2 = [target_sum / n for _ in range(n)]\n \n new_population.append(child1)\n new_population.append(child2)\n \n # Replace old population with new one\n population = new_population\n # Evaluate new population\n scores = [evaluate(seq) for seq in population]\n \n # Track the best sequence of the generation\n current_best_idx = np.argmin(scores)\n current_best_score = scores[current_best_idx]\n current_best_seq = population[current_best_idx].copy()\n \n if current_best_score < best_score:\n best_score = current_best_score\n best_seq = current_best_seq.copy()\n print(f\"New best score: {best_score}\")\n \n generations += 1\n \n return [float(x) for x in best_seq]\n```",
"env/all/time/policy": 328.0465613431297,
"env/all/time/policy/min": 140.11744737625122,
"env/all/time/policy/max": 457.1768820285797,
"env/all/time/env_step": 998.7551159900613,
"env/all/time/env_step/min": 0.5925302505493164,
"env/all/time/env_step/max": 2655.849058866501,
"env/all/time/reward_compute": 4.256144165992737e-07,
"env/all/time/reward_compute/min": 2.0116567611694336e-07,
"env/all/time/reward_compute/max": 1.1771917343139648e-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": 0.01119990460574627,
"advantage/min": -1.0,
"advantage/max": 2.0784976482391357,
"time/assemble_training_data": 10.248331785202026,
"time/kl_vs_base": 101.55124521255493,
"kl_policy_base": 0.0007331730448640883,
"time/train": 560.2120199203491,
"time/save_checkpoint": 35.84979319572449,
"time/total": 3814.2435314655304
}
[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