# Full 500-task evaluation with five self-evolution rounds; 16 candidates per train/eval task.
seed: 43
data_seed: 42
model:
  name: Qwen/Qwen2.5-Coder-1.5B-Instruct
  device: auto
  dtype: bfloat16
  attn_implementation: sdpa
data:
  train: data/mbpp/train.jsonl
  calibration: data/mbpp/calibration.jsonl
  validation: data/mbpp/validation.jsonl
  eval: data/mbpp/eval.jsonl
output_dir: runs/retention_5round_single_seed_train16_eval16_b128_v1
data_limits: {}
methods:
- plain
- spd_hard
- spectral_soft
rounds: 5
checkpoint_retention: latest
generation:
  train_samples: 16
  eval_samples: 16
  batch_size: 16
  max_new_tokens: 512
  max_prompt_tokens: 1024
  temperature: 0.8
  top_p: 0.95
  top_k: 0
  task_batch_size: 128
  sequence_batch_size: 128
calibration:
  max_examples: 50
  max_length: 1536
  span_mode: completion
  layers: null
  rank_fraction: 0.5
  tau: 1.0
  rho: 0.5
train:
  epochs: 1
  batch_size: 1
  gradient_accumulation_steps: 16
  max_length: 1536
  learning_rate: 1.0e-05
  weight_decay: 0.01
  warmup_ratio: 0.03
  max_grad_norm: 1.0
  lora_rank: 8
  lora_alpha: 8
  lora_dropout: 0.05
  gradient_checkpointing: true
  loss_scope: all
evaluation:
  backend: local
  allow_unsafe_local: true
  code_extraction: first_fence
  workers: 8
  timeout: 5
  ks:
  - 1
  - 8
  - 16
  correct_budget: 4
  correct_budgets:
  - 4
  - 8
  - 16
  correctness_margin: 0.01
  bootstrap_samples: 2000
diagnostics:
  evaluate_generation_policy: false
  eval_task_limit: 128
  eval_samples: 16