============================================================
Refinement started at 2026-06-25 08:29:29
============================================================
Logging to ./refined_adapters/shell/refine.log
Args: Namespace(language='shell', results_dir='ankhanhtran02/executed_calibration_results', results_source='hf_hub', results_repo_type='model', adapter_path='ankhanhtran02/lora-per-task-executable-start-4', model_name_or_path='Qwen/Qwen2.5-Coder-1.5B', output_dir='./refined_adapters/shell', per_device_train_batch_size=1, num_train_epochs=3, learning_rate=5e-05, weight_decay=0.01, gradient_accumulation_steps=8, num_warmup_steps=0, max_prompt_len=4096, max_ans_len=2048, seed=42, logging_steps=5, max_samples=600, local_rank=0, offload=False, zero_stage=0, gradient_checkpointing=False, disable_dropout=False, run_name='refine_shell', group_name='refine_adapter', enable_wandb=False, enable_tensorboard=False, tensorboard_path='refine_tensorboard', deepspeed=True, deepspeed_config=None, deepscale=False, deepscale_config=None, global_rank=0)
[refine] Loaded adapter from HF Hub: ankhanhtran02/lora-per-task-executable-start-4/shell/0
trainable params: 1,867,776 || all params: 1,545,176,576 || trainable%: 0.1208778355180036
[repair] extracted 19 pairs (skipped: 162 no-fail-traceback, 201 no-label)
[refine] 19 repair pairs for language=shell
Time to load fused_adam op: 1.1099557876586914 seconds
***** Running adapter refinement *****
  language      = shell
  train pairs   = 19
  epochs        = 3
  steps/epoch   = 5
  total steps   = 15
  batch/device  = 1
  grad accum    = 8
  lr            = 5e-05
[epoch 1/3] step 5/5 global_step 5 loss 0.3172
[refine] Epoch 1/3 complete.
[epoch 2/3] step 5/5 global_step 10 loss 0.2476
[refine] Epoch 2/3 complete.
[epoch 3/3] step 5/5 global_step 15 loss 0.2535
[refine] Epoch 3/3 complete.
[refine] Saved fine-tuned adapter to ./refined_adapters/shell