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

Training Loss	Epoch	Step	Validation Loss
14.2348	0.0001	1	14.2931
2.9296	0.0414	500	3.0717
2.6209	0.0829	1000	2.7409
2.4945	0.1243	1500	2.6571
2.5314	0.1657	2000	2.6148
2.4494	0.2072	2500	2.5855
2.4257	0.2486	3000	2.5569
2.4346	0.2901	3500	2.5402
2.4243	0.3315	4000	2.5284
2.3699	0.3729	4500	2.5134
2.4072	0.4144	5000	2.5064
2.4064	0.4558	5500	2.4953
2.3683	0.4972	6000	2.4950
2.3811	0.5387	6500	2.4804
2.3494	0.5801	7000	2.4734
2.3444	0.6215	7500	2.4663
2.3578	0.6630	8000	2.4626
2.344	0.7044	8500	2.4557
2.3284	0.7458	9000	2.4529
2.3086	0.7873	9500	2.4488
2.321	0.8287	10000	2.4443
2.3339	0.8702	10500	2.4409
2.3236	0.9116	11000	2.4391
2.3404	0.9530	11500	2.4387
2.3103	0.9945	12000	2.4342

Framework versions

Transformers 4.46.3
Pytorch 2.5.1+cu124
Datasets 3.1.0
Tokenizers 0.20.3