Downloads · 30 days
0
EdwardoSunny/ttt-discover-v3-qwen3-8b-ac1-s2-step14
ttt-discover-v3-qwen3-8b-ac1-s2-step14 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 13 (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 13 (0-indexed). Strict upstream eval parity:
1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.
{
"step": 13,
"progress/batch": 13,
"optim/lr": 4e-05,
"progress/done_frac": 0.28,
"puct/buffer_size": 216,
"puct/sampled_size": 8,
"puct/T": 6656,
"puct/scale_last": 0.5734890404777317,
"puct/buffer_value/mean": -1.554290119725878,
"puct/buffer_value/std": 0.12970209707276134,
"puct/buffer_value/min": -2.0797874683482562,
"puct/buffer_value/max": -1.5062984278705245,
"puct/buffer_timestep/mean": 5.7407407407407405,
"puct/buffer_timestep/std": 3.9024455067027692,
"puct/buffer_timestep/min": -1.0,
"puct/buffer_timestep/max": 12.0,
"puct/buffer_construction_len/mean": 1277.7314814814815,
"puct/buffer_construction_len/std": 1100.7163375164298,
"puct/buffer_construction_len/min": 1000.0,
"puct/buffer_construction_len/max": 7850.0,
"puct/sampled_value/mean": -1.50630303394062,
"puct/sampled_value/std": 1.7409308562693965e-06,
"puct/sampled_value/min": -1.5063036919506336,
"puct/sampled_value/max": -1.5062984278705245,
"puct/sampled_timestep/mean": 12.0,
"puct/sampled_timestep/std": 0.0,
"puct/sampled_timestep/min": 12.0,
"puct/sampled_timestep/max": 12.0,
"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": 5228.581015110016,
"env/all/ac_tokens_per_turn": 8265.84375,
"env/all/ob_tokens_per_turn": 2756.5,
"env/all/turns_per_episode": 1.0,
"env/all/total_episodes": 512,
"env/all/total_turns": 512,
"env/all/total_ac_tokens": 4232112,
"env/all/total_ob_tokens": 1411328,
"env/all/time/sampling_mean": 278.4598410818726,
"env/all/time/sampling_max": 383.5143976211548,
"env/all/time/env_step_mean": 2441.4981209593825,
"env/all/time/env_step_max": 4867.755452156067,
"env/all/reward/mean": 0.5224474789857925,
"env/all/reward/max": 0.6639215197918972,
"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": 0.5224474789857925,
"env/all/correctness": 0.84765625,
"env/all/correctness/min": 0.0,
"env/all/correctness/max": 1.0,
"env/all/raw_score": 2.1199714197592625,
"env/all/raw_score/min": 1.5062021090598625,
"env/all/raw_score/max": 86.02816382786607,
"env/all/initial_raw_score": -1.5063030339406203,
"env/all/initial_raw_score/min": -1.5063036919506336,
"env/all/initial_raw_score/max": -1.5062984278705245,
"env/all/msg": "RuntimeError: Program execution failed: IndexError: Cannot choose from an empty sequence",
"env/all/parsed_code": "```python\nimport time\nimport numpy as np\nimport random\n\ndef propose_candidate(seed=42, **kwargs):\n \"\"\"\n Genetic Algorithm for Optimal Sequence Search\n\n This method uses a population-based approach to evolve sequences of non-negative\n coefficients for a step function to minimize the evaluation score defined by\n the provided evaluate_sequence function. The algorithm maintains diversity \n through selection, crossover, and mutation, aiming to explore the search space \n more effectively than local perturbation methods.\n\n Key Strategies:\n - Maintains a population of sequences.\n - Uses tournament selection and crossover with mutation to generate offspring.\n - Balances exploration and exploitation by varying mutation rates.\n - Leverages population diversity to avoid local minima.\n - Adapts mutation scale based on current population statistics.\n \"\"\"\n np.random.seed(seed)\n deadline = time.time() + 1000 - 10\n\n # Initialize population with diverse starting points\n population_size = 50\n population = []\n if 'height_sequence_1' in globals() and isinstance(height_sequence_1, (list, np.ndarray)):\n # Add the previous best sequence to the population\n best_prev = list(np.asarray(height_sequence_1, dtype=float))\n population.append(best_prev)\n else:\n # Generate diverse initial sequences\n for _ in range(population_size):\n if np.random.rand() < 0.4:\n base = np.random.uniform(0.01, 0.05, size=1000)\n elif np.random.rand() < 0.5:\n base = np.random.uniform(0.0001, 0.001, size=1000)\n else:\n base = np.random.exponential(scale=0.01, size=1000)\n sum_base = np.sum(base)\n scale_factor = 0.1 / sum_base\n seq = [max(0.0, x * scale_factor) for x in base]\n population.append(seq)\n\n best_sequence = None\n best_score = float('inf')\n\n while time.time() < deadline:\n # Evaluate all sequences to get fitness\n evaluated = []\n for seq in population:\n try:\n score = evaluate_sequence(seq)\n if score != np.inf:\n evaluated.append((score, seq))\n except:\n # Skip invalid sequences\n pass\n\n # If no valid sequences, break early\n if not evaluated:\n break\n\n # Sort by evaluation score (ascending)\n evaluated.sort()\n evaluated = evaluated[:int(len(evaluated)*0.8)] # Keep top 80%\n\n # Select parents\n num_parents = int(len(evaluated) * 0.2)\n parents = [seq for score, seq in evaluated[:num_parents]]\n\n # Generate new population\n new_population = []\n new_population.extend(parents)\n\n # Generate offspring\n while len(new_population) < population_size:\n # Select parents\n parent1 = random.choice(parents)\n parent2 = random.choice(parents)\n\n # Crossover\n child = []\n for i in range(len(parent1)):\n if random.random() < 0.5:\n child.append(parent1[i])\n else:\n child.append(parent2[i])\n\n # Mutate\n for i in range(len(child)):\n if random.random() < 0.1:\n scale = 0.2 * np.std(child)\n mutation = np.random.exponential(scale)\n new_val = max(0.0, child[i] + mutation)\n child[i] = new_val\n\n new_population.append(child)\n\n # Replace old population\n population = new_population\n\n # Update best sequence\n for score, seq in evaluated:\n if score < best_score:\n best_score = score\n best_sequence = seq\n\n # Evaluate new sequences\n for seq in population:\n try:\n score = evaluate_sequence(seq)\n if score < best_score:\n best_score = score\n best_sequence = seq\n except:\n pass\n\n # Return the best sequence found\n return [float(max(0.0, x)) for x in best_sequence]\n```",
"env/all/time/policy": 278.4598410818726,
"env/all/time/policy/min": 132.06126618385315,
"env/all/time/policy/max": 383.5143976211548,
"env/all/time/env_step": 2441.4981209593825,
"env/all/time/env_step/min": 0.00864100456237793,
"env/all/time/env_step/max": 4867.755452156067,
"env/all/time/reward_compute": 1.0314397513866425e-06,
"env/all/time/reward_compute/min": 1.862645149230957e-07,
"env/all/time/reward_compute/max": 2.8386712074279785e-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": 0.01855834573507309,
"advantage/min": -1.0,
"advantage/max": 2.1308395862579346,
"time/assemble_training_data": 9.406537532806396,
"time/kl_vs_base": 89.97477412223816,
"kl_policy_base": 0.0008465369464829564,
"time/train": 522.8746664524078,
"time/save_checkpoint": 18.338593244552612,
"time/total": 5871.459002494812
}
[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