See axolotl config

axolotl version: 0.8.0

yaml

1base_model: Dans-DiscountModels/7b-m-dans-personalityengine-v1.2.1-rc-2 2model_type: AutoModelForCausalLM 3tokenizer_type: AutoTokenizer 4 5trust_remote_code: 6 7# wandb configuration 8wandb_project: 7b-m-dans-optimizersweeps 9wandb_watch: 10 11wandb_run_id: repremover-1-1-ademamix_8bit-hi-lr-b1_0.9-b2_0.999-b3_0.999-a5 12wandb_log_model: 13 14# push checkpoints to hub 15hub_model_id: Dans-DiscountModels/7b-m-dans-optimizersweeps-repremover-1-ademamix_8bit-hi-lr-b1_0.9-b2_0.999-b3_0.999-a5 16# how to push checkpoints to hub 17# https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy 18hub_strategy: "every_save" 19# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets 20# Required to be true when used in combination with `push_dataset_to_hub` 21hf_use_auth_token: true 22 23# where to save the finished model to 24output_dir: ./7b-m-dans-optimizersweeps 25 26# where to save the dataset to 27dataset_prepared_path: ./7b-m-dans-optimizersweeps-data 28 29save_safetensors: true 30 31# dataset settings (local or huggingface repo) 32datasets: 33 - path: Dans-DiscountModels/pretokenization-test-3 34 ds_type: parquet 35 type: 36 37plugins: 38 - axolotl.integrations.liger.LigerPlugin 39 - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin 40liger_rope: true 41liger_rms_norm: true 42liger_layer_norm: true 43liger_glu_activation: true 44liger_fused_linear_cross_entropy: false 45cut_cross_entropy: true 46 47load_in_8bit: false 48load_in_4bit: false 49strict: false 50 51adapter: 52lora_model_dir: 53 54val_set_size: 0.01 55sequence_len: 8192 56 57sample_packing: false 58eval_sample_packing: false 59 60pad_to_sequence_len: true 61 62gradient_checkpointing: true 63# gradient_checkpointing_kwargs: 64# use_reentrant: false 65 66gradient_accumulation_steps: 1 67micro_batch_size: 4 68 69num_epochs: 3 70 71optimizer: ademamix_8bit 72optim_args: "beta1=0.9,beta2=0.999,beta3=0.999,alpha=5" 73 74lr_scheduler: rex 75learning_rate: 0.0000003 76cosine_min_lr_ratio: 77 78# weight_decay: 0.03 79max_grad_norm: 0.001 80 81train_on_inputs: false 82group_by_length: true 83 84bf16: true 85fp16: false 86tf32: false 87 88early_stopping_patience: 89 90resume_from_checkpoint: 91auto_resume_from_checkpoints: false 92 93local_rank: 94logging_steps: 1 95xformers_attention: 96flash_attention: true 97 98warmup_ratio: 0.1 99 100evals_per_epoch: 24 101eval_table_size: 102eval_max_new_tokens: 103 104saves_per_epoch: 1 105save_total_limit: 2 106 107debug: false 108 109deepspeed: deepspeed_configs/zero3_bf16.json 110 111fsdp: 112fsdp_config: 113 114special_tokens: 115

7b-m-dans-optimizersweeps-repremover-1-ademamix_8bit-hi-lr-b1_0.9-b2_0.999-b3_0.999-a5

• Loss: 2.0806

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• learning_rate: 3e-07
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 32
• total_eval_batch_size: 32
• optimizer: Use ademamix_8bit and the args are: beta1=0.9,beta2=0.999,beta3=0.999,alpha=5
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 41
• num_epochs: 3.0

Training results

Framework versions

• Transformers 4.51.3
• Pytorch 2.5.1+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1