| adam_beta1: 0.9 | |
| adam_beta2: 0.95 | |
| adjust_step: 0 | |
| autoresume: false | |
| batch_size: 6 | |
| clip_grad_norm: 1.0 | |
| comment: null | |
| cycle_length: null | |
| dtype: bfloat16 | |
| emb_freeze: null | |
| eval_dataset_path: /work01/yanokazuki/fineweb/valid_data_gpt2/ | |
| eval_every: 1000 | |
| first_attention: false | |
| first_attention_resume: false | |
| first_layer: true | |
| gradient_accumulation: 13 | |
| keep_checkpoints: null | |
| layer_freeze: null | |
| layer_freeze_2: false | |
| load_optimizer_state_on_resume: true | |
| lr: 0.0004 | |
| max_length: 1024 | |
| max_train_tokens: null | |
| min_lr_ratio: 0.1 | |
| model_config: model_config/478m.json | |
| model_name_or_path: null | |
| model_revision: null | |
| num_training_steps: 15000 | |
| optimizer: Adam | |
| restart_warmup_steps: null | |
| resume_from: null | |
| run_name: first_layer | |
| save_dir: checkpoints/first_layer | |
| save_every: 1000 | |
| scheduler: cosine | |
| seed: 0 | |
| shuffle: true | |
| skip_batches: !!set {} | |
| tags: | |
| - 396m-for-680m | |
| total_batch_size: 624 | |
| train_dataset_path: /work01/yanokazuki/fineweb/train_data_gpt2/ | |
| training_config: training_config/two_stage/478m_first_layer.yaml | |
| wandb_watch: true | |
| warmed_up_model: null | |
| warmup_steps: 1500 | |
| weight_decay: 0.0 | |
| workers: 8 | |