# ====== Model Configuration ======
base_model: Qwen/Qwen3-1.7B-Base
load_in_8bit: false
load_in_4bit: true
strict: false

# LoRA
adapter: qlora
lora_r: 32
lora_alpha: 32
lora_target_linear: true
lora_qkv_kernel: true
lora_o_kernel: true
lora_mlp_kernel: true
embeddings_skip_upcast: true

# Integration
xformers_attention: true
plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_layer_norm: true
liger_fused_linear_cross_entropy: true
flash_attention: false

# DeepSpeed
# deepspeed: "/kaggle/working/axolotl/deepspeed_configs/zero2_torch_compile.json"

# ====== Hyperparameter Configuration ======
sample_packing: true
learning_rate: 1.8e-4
sequence_len: 4096
micro_batch_size: 8
gradient_accumulation_steps: 1
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false

optimizer: adamw_8bit
lr_scheduler: cosine_with_min_lr
lr_scheduler_kwargs: {"min_lr_rate": 0.1}

warmup_ratio: 0.03
weight_decay: 0.01
fp16: true

max_grad_norm: 1.0
num_epochs: 1
save_total_limit: 2
saves_per_epoch: 1
logging_steps: 1

output_dir: /kaggle/working/outputs/qwen-sft-mt-kmvi
chat_template: qwen3

# Dataset
datasets:
  - path: Tung177/km-vi-translation
    type: chat_template
    split: "train"
    roles_to_train: ["assistant"]
dataset_prepared_path: last_run_prepared

dataloader_prefetch_factor: 8
dataloader_num_workers: 2
dataloader_pin_memory: true

# ====== Tracking ======
wandb_project: MT-V1
wandb_name: qwen3-1.7B-mt-kmvi-v0.2
wandb_log_model: "false"

kaggle/working/outputs/qwen-sft-mt-kmvi

This model is a fine-tuned version of Qwen/Qwen3-1.7B-Base on the Tung177/km-vi-translation 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: 0.00018
train_batch_size: 8
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 2
total_train_batch_size: 16
total_eval_batch_size: 16
optimizer: Use OptimizerNames.ADAMW_8BIT with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine_with_min_lr
lr_scheduler_warmup_steps: 36
training_steps: 1230
mixed_precision_training: Native AMP

Training results

Framework versions

PEFT 0.17.1
Transformers 4.57.1
Pytorch 2.6.0+cu124
Datasets 4.0.0
Tokenizers 0.22.1

Tung177/Qwen3-1.7B-KmVi-500K-v0.2

kaggle/working/outputs/qwen-sft-mt-kmvi

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

Tung177/
Qwen3-1.7B-KmVi-500K-v0.2