# Training config for the General-tier adapter.
# Copy this file per premium model and change base_model / output_dir / hub_repo.

base_model: Qwen/Qwen2.5-7B-Instruct
dataset_path: data/finance_sft          # from src.data.prepare_datasets (load_from_disk)
# dataset_hub: finpy1789/finance-sft-mix  # or load from the Hub instead
output_dir: outputs/general
hub_repo: finpy1789/finllm-general-qwen2.5-7b

method: qlora            # lora | qlora | dora | adalora
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
target_modules: all-linear

max_length: 1024
num_train_epochs: 2
per_device_train_batch_size: 4
gradient_accumulation_steps: 4     # effective batch 16
learning_rate: 1.0e-4
lr_scheduler_type: cosine
warmup_ratio: 0.03
weight_decay: 0.001
max_grad_norm: 0.3
optim: paged_adamw_32bit
logging_steps: 25
save_steps: 250
eval_steps: 250
packing: false
seed: 42

# preference stage (src.training.preference)
pref_dataset: argilla/ultrafeedback-binarized-preferences-cleaned  # placeholder until
                                                                   # finance preference pairs exist
pref_learning_rate: 5.0e-6
pref_beta: 0.1
pref_max_steps: 500