### Qwen3-1.7B Full SFT Config (H200x4 Optimized)
model_name_or_path: models/Qwen3-1.7B
stage: sft
do_train: true
finetuning_type: full

dataset: preference_extractor_train
template: qwen
cutoff_len: 1024
overwrite_cache: true
preprocessing_num_workers: 4

output_dir: saves/qwen3-1.7b-full-sft-h200
logging_steps: 5
save_strategy: steps
save_steps: 200
plot_loss: true
overwrite_output_dir: true

# H200x4 Configuration
# Total Batch Size = 32 * 4 * 1 = 128
per_device_train_batch_size: 32
gradient_accumulation_steps: 1
learning_rate: 2.0e-5
num_train_epochs: 1.0
lr_scheduler_type: cosine
warmup_ratio: 0.05
bf16: true
flash_attn: fa2

val_size: 0.01
per_device_eval_batch_size: 32
eval_strategy: steps
eval_steps: 200