adapter: lora
base_model: Qwen/Qwen3-8B
bf16: true
bnb_4bit_compute_dtype: bfloat16
bnb_4bit_quant_type: nf4
bnb_4bit_use_double_quant: true
dataset_prepared_path: out/prepared_dataset_persistent_seed777
datasets:
- field_messages: messages
  message_field_content: content
  message_field_role: role
  path: /e/project1/reformo/nguyen38/pythonformer-workshop/train/out/paired_data/persistent/traces.jsonl
  roles_to_train:
  - assistant
  type: chat_template
eval_steps: 5
flash_attention: true
gradient_accumulation_steps: 16
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
learning_rate: 0.0001
load_in_4bit: true
load_in_8bit: false
logging_steps: 1
lora_alpha: 128
lora_dropout: 0.05
lora_r: 64
lora_target_linear: false
lora_target_modules:
- q_proj
- k_proj
- v_proj
- o_proj
- gate_proj
- up_proj
- down_proj
lr_scheduler: cosine
micro_batch_size: 1
model_type: AutoModelForCausalLM
num_epochs: 3.0
optimizer: adamw_torch
output_dir: out/qwen3-8b-persistent-seed777
pad_to_sequence_len: true
sample_packing: false
save_strategy: epoch
save_total_limit: 3
seed: 777
sequence_len: 16384
strict: false
tf32: true
tokenizer_type: AutoTokenizer
trust_remote_code: true
val_set_size: 0.04
wandb_log_model: null
wandb_watch: null
warmup_ratio: 0.03
weight_decay: 0.01