| # 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 | |