training_config.yaml

data:
  collator:
    pad_to_multiple_of: 8
  dataloader:
    drop_last: true
    num_workers: 4
    pin_memory: true
    shuffle: true
  processed_dir: finetune_processed_experiences
fsdp:
  activation_checkpointing: true
  mixed_precision: true
  sharding_strategy: FULL_SHARD
gpu:
  data_parallel: true
  single_gpu: false
huggingface:
  create_model_card: true
  repo_name: dtadpole/KernelCoder-4B_20250621-071556
  upload: true
lora:
  alpha: 64
  bias: none
  dropout: 0.05
  r: 64
  target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - down_proj
  - up_proj
model:
  dtype: null
  load_in_4bit: true
  max_seq_length: 16384
  name: Qwen/Qwen3-4B
test:
  default_prompt: '<|im_start|>system

    You are a helpful assistant.<|im_end|>

    <|im_start|>user

    What is machine learning?<|im_end|>

    <|im_start|>assistant

    '
  generation:
    do_sample: true
    max_new_tokens: 1024
    temperature: 0.7
    use_cache: true
training:
  gradient_accumulation_steps: 1
  learning_rate: 5.0e-05
  logging_steps: 1
  lr_scheduler_type: cosine
  max_grad_norm: 0.75
  max_steps: -1
  num_train_epochs: 2
  num_workers: 4
  optim: paged_adamw_8bit
  output_dir: ../finetune_model_output
  per_device_batch_size: 1
  save_steps: 100
  save_total_limit: 3
  seed: 3407
  use_custom_loss_masking: true
  warmup_steps: 10
  weight_decay: 0.05