base_model: google/gemma-4-31B-it
plugins:
  - axolotl.integrations.liger.LigerPlugin
cut_cross_entropy: false
torch_compile: false
liger_layer_norm: true
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_rms_norm_gated: true
strict: false
chat_template: gemma4
chat_template_kwargs:
  enable_thinking: false

sequence_len: 2048
sample_packing: false
pad_to_sequence_len: false
dataset_num_proc: 2
adapter: lora

lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules: 'model.language_model.layers.[\d]+.(_checkpoint_wrapped_module.)?(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'

micro_batch_size: 1
gradient_accumulation_steps: 1
num_epochs: 1
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 5.0e-5
warmup_ratio: 0.05

bf16: true
tf32: true
attn_implementation: sdpa
gradient_checkpointing: false
weight_decay: 0.0

logging_steps: 10
evals_per_epoch: 2
saves_per_epoch: 1

fsdp_version: 2
fsdp_config:
  offload_params: false
  state_dict_type: SHARDED_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: Gemma4TextDecoderLayer
  reshard_after_forward: true
  cpu_ram_efficient_loading: true

dp_shard_size: 4
experimental_skip_move_to_device: true

use_wandb: true
wandb_project: korail-gemma4
wandb_name: korail-gemma4-focus-sft-v5

push_to_hub: false

mnt/cepheid/users/hsypfsv/KORAIL_MAINTENANCE/finetune/gemma4_31b/output/korail_gemma4_h100_lora_4gpu_focus_sft_v5

This model is a fine-tuned version of google/gemma-4-31B-it on the /mnt/cepheid/users/hsypfsv/KORAIL_MAINTENANCE/finetune/gemma4_31b/data/korail_focus_sft_train.jsonl dataset. It achieves the following results on the evaluation set:

Loss: 1.5416
Ppl: 4.6721
Memory/max Active (gib): 26.69
Memory/max Allocated (gib): 26.69
Memory/device Reserved (gib): 36.52

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: 5e-05
train_batch_size: 1
eval_batch_size: 1
seed: 42
distributed_type: multi-GPU
num_devices: 4
total_train_batch_size: 4
total_eval_batch_size: 4
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED 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: 33
training_steps: 675

Training results

Training Loss	Epoch	Step	Validation Loss	Ppl	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	6.6640	783.6891	24.22	24.22	24.51
1.5144	0.5007	338	1.5955	4.9309	27.33	27.33	38.51
1.6051	1.0	675	1.5416	4.6721	26.69	26.69	36.52

Framework versions

Transformers 5.9.0
Pytorch 2.12.0+cu130
Datasets 4.8.5- Tokenizers 0.22.2

hyunseop/RailWayGemma31B-LoRA

mnt/cepheid/users/hsypfsv/KORAIL_MAINTENANCE/finetune/gemma4_31b/output/korail_gemma4_h100_lora_4gpu_focus_sft_v5

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

hyunseop/
RailWayGemma31B-LoRA