| bf16: true | |
| cutoff_len: 2048 | |
| dataset: Markie_Voss_d10000 | |
| dataset_dir: /workspace/LlamaFactory/data | |
| ddp_timeout: 180000000 | |
| do_train: true | |
| do_eval: false | |
| enable_thinking: false | |
| finetuning_type: lora | |
| lora_alpha: 64 | |
| lora_rank: 32 | |
| lora_dropout: 0.03 | |
| lora_target: all | |
| flash_attn: auto | |
| gradient_accumulation_steps: 8 | |
| include_num_input_tokens_seen: true | |
| learning_rate: 1e-4 | |
| logging_steps: 1 | |
| lr_scheduler_type: cosine | |
| max_grad_norm: 1 | |
| max_samples: 100000000 | |
| model_name_or_path: /workspace/Qwen/Qwen3-8B-Base | |
| num_train_epochs: 10 | |
| optim: adamw_torch | |
| output_dir: /workspace/v127rc_exp2/B_rep | |
| packing: true | |
| per_device_train_batch_size: 1 | |
| plot_loss: true | |
| preprocessing_num_workers: 16 | |
| report_to: wandb | |
| save_steps: 100 | |
| save_only_model: true | |
| stage: pt | |
| template: qwen3_nothink | |
| trust_remote_code: true | |
| warmup_ratio: 0.01 | |
| weight_decay: 0.01 | |
| adam_beta1: 0.9 | |
| adam_beta2: 0.95 | |