Ttt Discover v3 Qwen3 8b Ac1 Step1

TTT-Discover clean run (v3) — ac1 — step 1

LoRA adapter (rank 32, all-linear) from the upstream-faithful local reproduction of TTT-Discover on autoresearch-bench ac1. Checkpoint saved after training step 0 (0-indexed). Strict upstream eval parity: 1100s hard kill, verbatim prompts/entrypoints, group 64x8, T=1.0, kl 0.1.

Step metrics

{
  "step": 0,
  "progress/batch": 0,
  "optim/lr": 4e-05,
  "progress/done_frac": 0.02,
  "puct/buffer_size": 8,
  "puct/sampled_size": 8,
  "puct/T": 0,
  "puct/scale_last": 1e-06,
  "puct/buffer_value/mean": -2.0000000000000027,
  "puct/buffer_value/std": 2.8651498730356417e-15,
  "puct/buffer_value/min": -2.000000000000008,
  "puct/buffer_value/max": -1.9999999999999987,
  "puct/buffer_timestep/mean": -1.0,
  "puct/buffer_timestep/std": 0.0,
  "puct/buffer_timestep/min": -1.0,
  "puct/buffer_timestep/max": -1.0,
  "puct/buffer_construction_len/mean": 4543.25,
  "puct/buffer_construction_len/std": 2163.362472518186,
  "puct/buffer_construction_len/min": 1493.0,
  "puct/buffer_construction_len/max": 7397.0,
  "puct/sampled_value/mean": -2.0000000000000027,
  "puct/sampled_value/std": 2.8651498730356417e-15,
  "puct/sampled_value/min": -2.000000000000008,
  "puct/sampled_value/max": -1.9999999999999987,
  "puct/sampled_timestep/mean": -1.0,
  "puct/sampled_timestep/std": 0.0,
  "puct/sampled_timestep/min": -1.0,
  "puct/sampled_timestep/max": -1.0,
  "puct/sampled_construction_len/mean": 4543.25,
  "puct/sampled_construction_len/std": 2163.362472518186,
  "puct/sampled_construction_len/min": 1493.0,
  "puct/sampled_construction_len/max": 7397.0,
  "time/sampling": 4370.824383497238,
  "env/all/ac_tokens_per_turn": 8505.33203125,
  "env/all/ob_tokens_per_turn": 2472.0,
  "env/all/turns_per_episode": 1.0,
  "env/all/total_episodes": 512,
  "env/all/total_turns": 512,
  "env/all/total_ac_tokens": 4354730,
  "env/all/total_ob_tokens": 1265664,
  "env/all/time/sampling_mean": 507.09334986144677,
  "env/all/time/sampling_max": 720.1925506591797,
  "env/all/time/env_step_mean": 1600.179204819724,
  "env/all/time/env_step_max": 3686.734853029251,
  "env/all/reward/mean": 0.36894933197170376,
  "env/all/reward/max": 0.6613869082722876,
  "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": 0.36894933197170376,
  "env/all/correctness": 0.693359375,
  "env/all/correctness/min": 0.0,
  "env/all/correctness/max": 1.0,
  "env/all/raw_score": 6.0977610238147415,
  "env/all/raw_score/min": 1.5119742784119927,
  "env/all/raw_score/max": 1277.243862113602,
  "env/all/initial_raw_score": -2.0000000000000027,
  "env/all/initial_raw_score/min": -2.000000000000008,
  "env/all/initial_raw_score/max": -1.9999999999999987,
  "env/all/msg": "RuntimeError: Program execution failed: ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()",
  "env/all/parsed_code": "```python\nimport numpy as np\nimport time\nimport random\n\ndef propose_candidate(seed=42, budget_s=1000, **kwargs):\n    \"\"\"\n    Implements a simulated annealing algorithm that starts from a given\n    sequence (height_sequence_1) and explores the search space to find a\n    sequence minimizing the evaluation function.\n    \"\"\"\n    # Access the global best sequence if available\n    global height_sequence_1\n    if 'height_sequence_1' in globals() and height_sequence_1:\n        initial_sequence = list(np.array(height_sequence_1, dtype=float))\n    else:\n        # Fallback: generate a sequence with random values between 0 and 1\n        initial_sequence = [np.random.rand() for _ in range(2003)]\n\n    current_sequence = initial_sequence.copy()\n    best_sequence = current_sequence.copy()\n    best_score = evaluate_sequence(best_sequence)\n\n    # Initialize annealing parameters\n    temperature = 100.0\n    cooling_rate = 0.995  # Slightly faster cooling\n    iterations = 0\n    start_time = time.time()\n\n    while time.time() - start_time < budget_s:\n        # Generate a random perturbation\n        idx = np.random.randint(len(current_sequence))\n        delta = random.uniform(-0.1, 0.1)  # Small random perturbation\n        new_value = max(0.0, min(1000.0, current_sequence[idx] + delta))\n        new_sequence = current_sequence.copy()\n        new_sequence[idx] = new_value\n\n        # Evaluate the new sequence\n        try:\n            new_score = evaluate_sequence(new_sequence)\n        except:\n            new_score = float('inf')\n\n        # Update best sequence if better, or accept with probability if worse\n        if new_score < best_score:\n            current_sequence = new_sequence\n            best_sequence = current_sequence.copy()\n            best_score = new_score\n        elif np.random.rand() < np.exp((best_score - new_score) / temperature):\n            current_sequence = new_sequence\n\n        # Cool down\n        temperature *= cooling_rate\n        iterations += 1\n\n    return [float(x) for x in best_sequence]\n```",
  "env/all/time/policy": 507.09334986144677,
  "env/all/time/policy/min": 69.81673073768616,
  "env/all/time/policy/max": 720.1925506591797,
  "env/all/time/env_step": 1600.179204819724,
  "env/all/time/env_step/min": 0.006195545196533203,
  "env/all/time/env_step/max": 3686.734853029251,
  "env/all/time/reward_compute": 8.731149137020111e-07,
  "env/all/time/reward_compute/min": 2.086162567138672e-07,
  "env/all/time/reward_compute/max": 2.5890767574310303e-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": 0.027576152235269547,
  "advantage/min": -0.9992027878761292,
  "advantage/max": 6.879091739654541,
  "time/assemble_training_data": 5.421263217926025,
  "time/kl_vs_base": 136.5565905570984,
  "kl_policy_base": 0.0004980333033017814,
  "time/train": 1030.374207496643,
  "time/save_checkpoint": 15.91596269607544,
  "time/total": 5563.602016210556
}

Hardware provenance

[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