Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step7
ttt-discover-v3-qwen3-8b-ac1-s2-step7 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 6 (0-indexed). Strict upstream eval parity: 1100s hard k…
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 ac1. Checkpoint saved
after training step 6 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 6,
"progress/batch": 6,
"optim/lr": 4e-05,
"progress/done_frac": 0.14,
"puct/buffer_size": 104,
"puct/sampled_size": 8,
"puct/T": 3072,
"puct/scale_last": 0.4917065830127707,
"puct/buffer_value/mean": -1.5875642895218651,
"puct/buffer_value/std": 0.16044655430586713,
"puct/buffer_value/min": -2.000000000000007,
"puct/buffer_value/max": -1.5065538969617303,
"puct/buffer_timestep/mean": 2.230769230769231,
"puct/buffer_timestep/std": 1.8873606379054302,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 5.0,
"puct/buffer_construction_len/mean": 1576.826923076923,
"puct/buffer_construction_len/std": 1530.9555176318665,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.5076437427443543,
"puct/sampled_value/std": 0.000624750105480906,
"puct/sampled_value/min": -1.5087257684146806,
"puct/sampled_value/max": -1.5065538969617303,
"puct/sampled_timestep/mean": 5.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 5.0,
"puct/sampled_timestep/max": 5.0,
"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": 5168.406282901764,
"env/all/ac_tokens_per_turn": 7674.57421875,
"env/all/ob_tokens_per_turn": 3058.625,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 3929382,
"env/all/total_ob_tokens": 1566016,
"env/all/time/sampling_mean": 259.0792518411763,
"env/all/time/sampling_max": 354.1538269519806,
"env/all/time/env_step_mean": 2141.49811778171,
"env/all/time/env_step_max": 4836.556226015091,
"env/all/reward/mean": 0.48488694167448354,
"env/all/reward/max": 0.6638767459075229,
"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": 0.48488694167448354,
"env/all/correctness": 0.779296875,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 1.9845716544252319,
"env/all/raw_score/min": 1.5063036919514743,
"env/all/raw_score/max": 129.41858526523762,
"env/all/initial_raw_score": -1.5076437427443543,
"env/all/initial_raw_score/min": -1.5087257684146806,
"env/all/initial_raw_score/max": -1.5065538969617303,
"env/all/msg": "RuntimeError: Program execution failed: ValueError: a must be 1-dimensional",
"env/all/parsed_code": "```python\n# EVOLVE-BLOCK-START\n\"\"\"Genetic Algorithm for Optimal Step Function Construction with Evolutionary Optimization.\"\"\"\nimport time\nimport numpy as np\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n np.random.seed(seed)\n deadline = time.time() + budget_s - 10\n n = 1000\n target_sum = np.sqrt(2 * n)\n \n # Initialize population with diversity\n population_size = 100\n population = []\n for _ in range(population_size):\n # Generate random sequence with clamping and normalization\n seq = np.random.rand(n) * 1000.0\n seq = np.clip(seq, 0.0, 1000.0)\n seq = seq / np.sum(seq) * target_sum\n population.append(seq)\n \n best_sequence = None\n best_score = float('inf')\n eval_count = 0\n \n while time.time() < deadline and len(population) > 0:\n # Evaluate all sequences\n fitness = []\n for seq in population:\n conv = np.convolve(seq, seq)\n current_score = np.max(conv)\n fitness.append((current_score, seq))\n \n # Update best solution\n if fitness:\n fitness.sort()\n curr_score = fitness[0][0]\n if curr_score < best_score:\n best_score = curr_score\n best_sequence = fitness[0][1]\n eval_count += 1\n print(f\"New best: {best_score}\")\n \n # Selection: top 50% as parents\n num_parents = int(len(population) / 2) if len(population) > 2 else 1\n parents = [fitness[i][1] for i in range(num_parents)]\n \n # Create offspring\n offspring = []\n while len(offspring) < len(population):\n # Crossover\n parent1, parent2 = np.random.choice(parents, 2, replace=False)\n child = (parent1 + parent2) / 2.0 # Blend crossover\n \n # Mutation: with 10% probability\n if np.random.rand() < 0.1:\n idx = np.random.randint(n)\n delta = np.random.normal(0, 0.05) * (1000.0 if np.random.rand() < 0.7 else 100.0)\n child[idx] += delta\n child[idx] = np.clip(child[idx], 0.0, 1000.0)\n \n # Normalize to target_sum\n sum_child = np.sum(child)\n if sum_child > 0:\n child = child / sum_child * target_sum\n else:\n child = np.full(n, target_sum / n)\n child = np.clip(child, 0.0, 1000.0)\n offspring.append(child)\n \n # Replace population\n population = offspring\n \n return [float(max(0.0, x)) for x in best_sequence]\n# EVOLVE-BLOCK-END\n```",
"env/all/time/policy": 259.0792518411763,
"env/all/time/policy/min": 122.45533061027527,
"env/all/time/policy/max": 354.1538269519806,
"env/all/time/env_step": 2141.49811778171,
"env/all/time/env_step/min": 0.0047566890716552734,
"env/all/time/env_step/max": 4836.556226015091,
"env/all/time/reward_compute": 5.1083043217659e-07,
"env/all/time/reward_compute/min": 2.2724270820617676e-07,
"env/all/time/reward_compute/max": 1.7471611499786377e-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": 0.02765469066798687,
"advantage/min": -1.0,
"advantage/max": 9.692758560180664,
"time/assemble_training_data": 9.020856380462646,
"time/kl_vs_base": 90.05235695838928,
"kl_policy_base": 0.0007476528408005834,
"time/train": 506.0403513908386,
"time/save_checkpoint": 22.575587034225464,
"time/total": 5797.858085155487
}
[2026-07-09T06:24:18+00:00] job=1812630 node=node-14 ngpu=3 ntrain=1 replicas=2 flash_attn=no
[2026-07-09T09:14:17+00:00] job=1813129 node=node-28 ngpu=3 ntrain=1 replicas=2 flash_attn=yes
[2026-07-09T13:46:36+00:00] job=1813130 node=node-22 ngpu=6 ntrain=2 replicas=4 flash_attn=yes