| add_bos: false | |
| checkpointing_steps: null | |
| clip_grad_norm: -1 | |
| config_name: null | |
| dataset_config_name: null | |
| dataset_name: null | |
| do_eval: true | |
| eval_batch_size: 2 | |
| eval_file: self-seq/data/lima500_withsys.jsonl | |
| eval_steps: '100' | |
| gradient_accumulation_steps: 16 | |
| gradient_checkpointing: true | |
| learning_rate: 2.0e-05 | |
| logging_steps: 5 | |
| lora_alpha: 16 | |
| lora_dropout: 0.1 | |
| lora_rank: 64 | |
| low_cpu_mem_usage: false | |
| lr_scheduler_type: linear | |
| max_seq_length: 2048 | |
| max_train_steps: 2316 | |
| model_name_or_path: /nas/shared/NLP_A100/model/Meta-Llama-3-8B | |
| model_revision: main | |
| num_train_epochs: 3 | |
| output_dir: output/self-seq-Meta-Llama-3-8B-tulu100k_base_ours_new_llama70b/ | |
| overwrite_cache: false | |
| per_device_train_batch_size: 2 | |
| preprocessing_num_workers: 64 | |
| prompt_template: tulu | |
| reduce_loss: mean | |
| report_to: all | |
| resume_from_checkpoint: null | |
| sample_train: -1 | |
| seed: 42 | |
| timeout: 1800 | |
| tokenizer_name: /nas/shared/NLP_A100/model/Meta-Llama-3-8B | |
| tokenizer_revision: null | |
| train_file: self-seq/data/tuluv2/tuluv2_100k_base_ours_new.jsonl | |
| trust_remote_code: false | |
| use_8bit_optimizer: false | |
| use_flash_attn: true | |
| use_lora: false | |
| use_qlora: false | |
| use_slow_tokenizer: true | |
| warmup_ratio: 0.03 | |
| weight_decay: 0.0 | |
| with_tracking: true | |