axolotl version: 0.5.2

base_model: meta-llama/Llama-3.1-8B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
tokenizer_use_fast: false
resize_token_embeddings_to_32x: false

flash_attention: true
xformers_attention:

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: skymizer/Llama3.1-base-tokenized-dolma-v1_7-50B
    train_on_split: train
    type: completion

test_datasets:
  - path: skymizer/Llama3.1-tokenized-dolma-v1_7-test
    split: test
    type: completion

is_preprocess: true
skip_prepare_dataset: true

dataset_prepared_path: /mnt/home/model-team/datasets/pretokenized/Llama3.1-8B-base-tokenized-dolma-v1_7_50B-4096 

hf_use_auth_token: true
output_dir: /mnt/home/model-team/models/Llama3.1-8B-relu-stage-2-dolma-50B-4096
resume_from_checkpoint:
auto_resume_from_checkpoints: true

sequence_len: 4096
sample_packing: true
sample_packing_group_size: 100000
sample_packing_bin_size: 200
pad_to_sequence_len: true

eval_sample_packing: false
# eval_causal_lm_metrics: ["perplexity"]

wandb_project: "sparse-tuning-cpt"
wandb_entity:
wandb_watch:
wandb_name: "Llama3.1-8B-relu-stage-1-dolma-50B-4096"
wandb_log_model:

# global batch size = 2 * 8 * 8 GPUs * 8 Nodes * 4096 = 4M
gradient_accumulation_steps: 8
micro_batch_size: 2
  # eval_batch_size: 2
num_epochs: 1
optimizer: adamw_torch
learning_rate: 0.000015
lr_scheduler: cosine
cosine_min_lr_ratio: 1.0 
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.95
adam_eps: 0.000001
max_grad_norm: 1.0

train_on_inputs: false
group_by_length: false
bf16: true
fp16:
tf32: false

hub_model_id: "skymizer/Llama3.1-8B-relu-stage-2-dolma-v1_7-50B-4096"

save_strategy: "steps"
save_steps: 500

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1

warmup_steps: 1
eval_steps: 500
eval_table_size:
debug:
deepspeed: /root/train/axolotl/deepspeed_configs/zero3_bf16.json
fsdp:
fsdp_config:
seed: 42

special_tokens:
  pad_token: "<|end_of_text|>"

Llama3.1-8B-relu-stage-2-dolma-v1_7-50B-4096

This model is a fine-tuned version of meta-llama/Llama-3.1-8B on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 2.4342

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1.5e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 64
• gradient_accumulation_steps: 8
• total_train_batch_size: 1024
• total_eval_batch_size: 128
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.95) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 2
• num_epochs: 1

Training results

Framework versions

• Transformers 4.46.3
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.20.3