See axolotl config

axolotl version: 0.12.1

base_model: aisingapore/Gemma-SEA-LION-v4-27B-IT
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

skip_prepare_dataset: false
remove_unused_columns: false
sample_packing: false

ddp_find_unused_parameters: true
deepspeed: deepspeed_configs/zero3.json

chat_template: gemma3
eot_tokens:
  - <end_of_turn>
datasets:
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_textcaps_EN_SEA
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xm3600_EN_SEA
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xflickr_EN_SEA_chunk_1
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xflickr_EN_SEA_chunk_2
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xflickr_EN_SEA_chunk_3
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xflickr_EN_SEA_chunk_4
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xflickr_EN_SEA_chunk_5
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xflickr_EN_SEA_chunk_6
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xflickr_EN_SEA_chunk_7
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xflickr_EN_SEA_chunk_8
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xflickr_EN_SEA_chunk_9
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_xflickr_EN_SEA_chunk_10
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_sea_vl_EN_SEA_chunk_1
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_sea_vl_EN_SEA_chunk_2
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_sea_vl_EN_SEA_chunk_3
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_sea_vl_EN_SEA_chunk_4
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_sea_vl_EN_SEA_chunk_5
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_sea_vl_EN_SEA_chunk_6
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_sea_vl_EN_SEA_chunk_7
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_sea_vl_EN_SEA_chunk_8
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_sea_vl_EN_SEA_chunk_9
    type: chat_template
    split: train
    field_messages: messages
  - path: /mnt/weka/all/holy/translation_parallel/pretrain_dataset_sea_vl_EN_SEA_chunk_10
    type: chat_template
    split: train
    field_messages: messages
dataset_prepared_path: peerat_test_path
val_set_size: 0.01
output_dir: ./outputs-pt/sealion-v4-gemma-3-27b-hero_pre_train_v2

sequence_len: 2048
pad_to_sequence_len: false

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 4
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 2e-5

bf16: true
fp16:
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
resume_from_checkpoint:
logging_steps: 1
flash_attention: false
# attn_implementation: sdpa

warmup_ratio: 0.1
evals_per_epoch: 1
save_steps: 1000
save_total_limit: 3
weight_decay: 0.0

# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

outputs-pt/sealion-v4-gemma-3-27b-hero_pre_train_v2

This model is a fine-tuned version of aisingapore/Gemma-SEA-LION-v4-27B-IT on the None dataset.

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: 2e-05
train_batch_size: 4
eval_batch_size: 4
seed: 42
distributed_type: multi-GPU
num_devices: 80
gradient_accumulation_steps: 2
total_train_batch_size: 640
total_eval_batch_size: 320
optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 1156
training_steps: 11568