tuandunghcmut/Qwen3-8B-Private

base_model: Qwen/Qwen3-8B
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name
hub_model_id: tuandunghcmut/Qwen3-8B-Private


plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
strict: false



chat_template: qwen3
datasets:
  # - path: trendmicro-ailab/Primus-Seed
  #   type: chat_template
    # split: train[:20%]
    # split: train
    # field_messages: conversations
    # message_property_mappings:
    #   role: from
    #   content: value
  - path: trendmicro-ailab/Primus-Reasoning
    type: chat_template
    # split: train[:20%]
    split: train
    split_thinking: true
    chat_template: qwen3
    field_messages: messages
    message_property_mappings:
      role: role
      content: content

      
val_set_size: 0.075
output_dir: ./outputs/out2
dataset_prepared_path: last_run_prepared

# sequence_len: 2048
sequence_len: 3072
sample_packing: true
eval_sample_packing: true


load_in_4bit: true
adapter: qlora
# lora_r: 16
# lora_alpha: 32

# lora_r: 32
# lora_alpha: 64


lora_r: 64
lora_alpha: 128

lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - down_proj
  - up_proj
lora_mlp_kernel: true
lora_qkv_kernel: true
lora_o_kernel: true

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 4
num_epochs: 30
optimizer: adamw_torch_4bit
lr_scheduler: cosine
learning_rate: 0.00002

bf16: auto
tf32: true

gradient_checkpointing: offload
gradient_checkpointing_kwargs:
  use_reentrant: false
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

warmup_ratio: 0.1
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.01
special_tokens:

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