# ===== Model =====

base_model: google/gemma-3-4b-it
processor_type: AutoProcessor

chat_template: gemma3

# 멀티모달(비전-챗) 필수 플래그
skip_prepare_dataset: true
remove_unused_columns: false
sample_packing: false

#shuffle_merged_datasets: false
#shuffle_before_merging_datasets: false   # (기본 false지만 명시 추천)

ddp_find_unused_parameters: true

dataloader_num_workers: 0

# ===== Data =====
eot_tokens:
  - <end_of_turn>
datasets:
  - path: vlm_data_2025101_1/gemma3-4b-v-KoV_2.0.1.jsonl
    type: chat_template
    field_messages: messages
    split: null

val_set_size: 0.0
dataset_prepared_path:

# ===== Output / Logging =====
output_dir: ./outputs/gemma3-4b-v-KoV_2.0.1_fft.jsonl
logging_steps: 1

# wandb 연동(원하면 변경/주석)
wandb_entity: minkyun1
wandb_project: kisti_vlm_axo
wandb_name: gemma3-4b-v-KoV_2.0.1_fft.jsonl

# ===== LoRA / Quantization =====
#adapter: lora
# LLaVA에서 언어모델 쪽 프로젝션에만 LoRA(안전 기본값)
#lora_r: 256
#lora_alpha: 512
#lora_dropout: 0.05
#lora_target_modules: "model.language_model.layers.[\\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj"

# 메모리 여유 충분하지만, 시작은 4bit 로 안정적으로
load_in_4bit: false
load_in_8bit: false
bf16: true
tf32: true
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
flash_attention: true
eager_attention:

# ===== Optim & Train =====
optimizer: adamw_torch_fused
learning_rate: 4e-5
lr_scheduler: cosine
warmup_ratio: 0.05
weight_decay: 0.01
max_grad_norm: 1.0
seed: 42
sequence_len: 8192 
pad_to_sequence_len: false
excess_length_strategy: drop

# GPU당 마이크로 배치/누적 → 유효 배치 = 1 * 8 * 2GPU = 16
micro_batch_size: 1
gradient_accumulation_steps: 16

num_epochs: 3
evals_per_epoch: 1
saves_per_epoch: 1
# save_first_step: true

# ===== Multi-GPU: DeepSpeed (추천) =====
# deepspeed 프리셋을 받아서 사용:
#   axolotl fetch deepspeed_configs
# 2×A100 80GB + 7B에는 zero2가 빠르고 안정적
deepspeed: ds_zero2.json

# ===== 디버그/재현성(선택) =====
# 데이터 전처리 멀티프로세스가 문제 생기면 1로 낮춰서 원인 파악
# dataset_processes: 1

# ===== [대안] FSDP2 설정(DeepSpeed 대신 쓰고 싶을 때) =====
# fsdp_version: 2
# fsdp_config:
#   offload_params: false
#   cpu_ram_efficient_loading: true
#   auto_wrap_policy: TRANSFORMER_BASED_WRAP
#   transformer_layer_cls_to_wrap: LlamaDecoderLayer
#   state_dict_type: FULL_STATE_DICT
#   reshard_after_forward: true

outputs/gemma3-4b-v-KoV_2.0.1_fft.jsonl

This model is a fine-tuned version of google/gemma-3-4b-it on the vlm_data_2025101_1/gemma3-4b-v-KoV_2.0.1.jsonl 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: 4e-05
train_batch_size: 1
eval_batch_size: 1
seed: 42
distributed_type: multi-GPU
num_devices: 8
gradient_accumulation_steps: 16
total_train_batch_size: 128
total_eval_batch_size: 8
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: 578
training_steps: 11567

Training results

Framework versions

Transformers 4.55.2
Pytorch 2.6.0+cu124
Datasets 4.0.0
Tokenizers 0.21.4

jjakgui/gemma3-4b-v-KoV_2.0.1_fft_20251027

outputs/gemma3-4b-v-KoV_2.0.1_fft.jsonl

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

jjakgui/
gemma3-4b-v-KoV_2.0.1_fft_20251027