| base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T | |
| model_type: LlamaForCausalLM | |
| tokenizer_type: AutoTokenizer | |
| load_in_8bit: false | |
| load_in_4bit: false | |
| strict: false | |
| pretraining_dataset: | |
| - path: augmxnt/shisa-pretrain-en-ja-v1 | |
| type: completion | |
| total_supervised_tokens: true | |
| pretrain_multipack_attn: false | |
| dataset_processes: 32 | |
| val_set_size: 0.0 | |
| output_dir: ./out | |
| pretrain_multipack_buffer_size: 100000 | |
| max_steps: 4702818 | |
| sequence_len: 2048 | |
| sample_packing: true | |
| pad_to_sequence_len: true | |
| eval_sample_packing: false | |
| adapter: | |
| lora_model_dir: | |
| lora_r: | |
| lora_alpha: | |
| lora_dropout: | |
| lora_target_linear: | |
| lora_fan_in_fan_out: | |
| wandb_project: tiny-llama | |
| wandb_entity: | |
| wandb_watch: | |
| wandb_name: | |
| wandb_log_model: | |
| gradient_accumulation_steps: 64 | |
| micro_batch_size: 1 | |
| num_epochs: 1 | |
| optimizer: adamw_apex_fused | |
| lr_scheduler: cosine | |
| learning_rate: 5e-5 | |
| adam_beta1: 0.9 | |
| adam_beta2: 0.95 | |
| train_on_inputs: false | |
| group_by_length: false | |
| bf16: true | |
| fp16: false | |
| tf32: false | |
| gradient_checkpointing: false | |
| early_stopping_patience: | |
| resume_from_checkpoint: | |
| auto_resume_from_checkpoints: | |
| logging_steps: 1 | |
| xformers_attention: | |
| flash_attention: true | |
| flash_attn_cross_entropy: false | |
| flash_attn_rms_norm: true | |
| flash_attn_fuse_qkv: false | |
| flash_attn_fuse_mlp: true | |
| save_total_limit: 15 | |
| warmup_steps: 100 | |
| evals_per_epoch: | |
| eval_table_size: | |
| save_steps: 250 | |
| saves_per_epoch: | |
| debug: | |
| deepspeed: deepspeed_configs/zero1.json | |
| weight_decay: 0.1 | |
| fsdp: | |
| fsdp_config: | |
| special_tokens: | |