| apollo_rank: 256 | |
| apollo_scale: 1 | |
| apollo_target: all | |
| apollo_update_interval: 200 | |
| bf16: true | |
| cutoff_len: 4096 | |
| dataset: codes_330k_ns | |
| dataset_dir: data | |
| ddp_timeout: 180000000 | |
| do_train: true | |
| enable_liger_kernel: true | |
| finetuning_type: freeze | |
| flash_attn: auto | |
| freeze_trainable_layers: 2 | |
| freeze_trainable_modules: all | |
| gradient_accumulation_steps: 8 | |
| include_num_input_tokens_seen: true | |
| learning_rate: 5.0e-05 | |
| logging_steps: 1 | |
| lr_scheduler_type: cosine | |
| max_grad_norm: 1.0 | |
| max_samples: 50000000 | |
| model_name_or_path: Qwen/Qwen2.5-Coder-7B-Instruct | |
| neat_packing: true | |
| num_train_epochs: 1.0 | |
| output_dir: saves/Qwen2.5-Coder-7B-Instruct/freeze/qwen_ns | |
| packing: true | |
| per_device_train_batch_size: 16 | |
| plot_loss: true | |
| preprocessing_num_workers: 16 | |
| report_to: none | |
| rope_scaling: llama3 | |
| save_steps: 500 | |
| stage: sft | |
| template: qwen | |
| trust_remote_code: true | |
| use_apollo: true | |
| use_llama_pro: true | |
| warmup_steps: 0 | |