• The 2e-4 and 5e-5 are definitely overtrained and perform significantly worse.
• 2e-5 is on the edge since weightwacher shows the embed as slightly overtrained for 2e-5, but NEFTune version is not
• 8e-6 performs the best, and 5e-6 also performed slightly better than 2e-5

See axolotl config

axolotl version: 0.4.0

yaml

1base_model: meta-llama/Meta-Llama-3-8B-Instruct 2model_type: LlamaForCausalLM 3tokenizer_type: AutoTokenizer 4 5load_in_8bit: false 6load_in_4bit: false 7strict: false 8 9chat_template: llama3 10datasets: 11 - path: augmxnt/ultra-orca-boros-en-ja-v1 12 type: sharegpt 13dataset_prepared_path: last_run_prepared 14val_set_size: 0.05 15output_dir: ./outputs/lr-5e6 16 17sequence_len: 8192 18sample_packing: true 19pad_to_sequence_len: true 20 21use_wandb: true 22wandb_project: shisa-v2 23wandb_entity: augmxnt 24wandb_name: shisa-v1-llama3-8b.lr-5e6 25 26gradient_accumulation_steps: 8 27micro_batch_size: 1 28num_epochs: 3 29optimizer: paged_adamw_8bit 30lr_scheduler: linear 31learning_rate: 5e-6 32 33train_on_inputs: false 34group_by_length: false 35bf16: auto 36fp16: 37tf32: false 38 39gradient_checkpointing: true 40gradient_checkpointing_kwargs: 41 use_reentrant: false 42early_stopping_patience: 43resume_from_checkpoint: 44logging_steps: 1 45xformers_attention: 46flash_attention: true 47 48warmup_steps: 100 49evals_per_epoch: 2 50eval_table_size: 51saves_per_epoch: 0 52debug: 53deepspeed: axolotl/deepspeed_configs/zero3_bf16.json 54weight_decay: 0.00 55fsdp: 56fsdp_config: 57special_tokens: 58 pad_token: <|end_of_text|> 59

outputs/lr-5e6

• Loss: 0.5020

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• learning_rate: 5e-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• num_epochs: 3

Training results

Framework versions

• Transformers 4.40.2
• Pytorch 2.3.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1