# Comparison run — 林路 persona LoRA on Qwen3.6-27B (dense VLM, text tower only).
# Dense model: standard attention everywhere, so the full q/k/v/o + MLP set is
# targetable (no fused-expert name clash).
# Run: CUDA_VISIBLE_DEVICES=1 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
#        llamafactory-cli train train_lora_27b.yaml

### model
model_name_or_path: ./Qwen3.6-27B
trust_remote_code: true

### method
stage: sft
do_train: true
finetuning_type: lora
lora_rank: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
freeze_vision_tower: true

### dataset
dataset: lin_lu_train
eval_dataset: lin_lu_val
dataset_dir: ./data
template: qwen
cutoff_len: 8192
overwrite_cache: true
preprocessing_num_workers: 8

### output
output_dir: ./out/lin_lu_lora_27b
logging_steps: 2
save_steps: 4          # ~10 min between checkpoints; host crashes lose little
save_total_limit: 3
plot_loss: true
overwrite_output_dir: true
report_to: none

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 1.0e-4
num_train_epochs: 5.0
lr_scheduler_type: cosine
warmup_ratio: 0.05
bf16: true
gradient_checkpointing: true

### eval
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 4
load_best_model_at_end: true
metric_for_best_model: eval_loss
greater_is_better: false