base_model_name: Qwen/Qwen2.5-1.5B-Instruct
max_seq_length: 512
max_new_tokens: 256
num_examples: 800
seed: 42
train_ratio: 0.8
val_ratio: 0.1
test_ratio: 0.1
task_type: general
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
- q_proj
- k_proj
- v_proj
- o_proj
- gate_proj
- up_proj
- down_proj
num_train_epochs: 6
per_device_train_batch_size: 1
per_device_eval_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 0.0001
weight_decay: 0.01
warmup_ratio: 0.05
logging_steps: 10
eval_steps: 40
save_steps: 40
save_total_limit: 2
early_stopping_patience: 4
gradient_checkpointing: true
dataloader_num_workers: 0
prefer_gpu: true
use_fp16: true
use_bf16: false
load_in_4bit: false
output_dir: outputs/promptforge-optimizer
dataset_path: data/promptforge_optimizer_dataset.csv
final_model_dir: outputs/promptforge-optimizer-model