qlora.yml

base_model: Qwen/Qwen1.5-14B
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

trust_remote_code: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: /data/data/final_set_cleaned/train/
    type: sharegpt
    conversation: chatml
    shards: 20
  - path: /data/data/map_coig_cqia.jsonl
    type: sharegpt
    conversation: chatml
  - path: /data/data/ruozhiba.jsonl
    type: sharegpt
    conversation: chatml
dataset_prepared_path:
val_set_size: 0 # 一定要写
output_dir: /data/qwen1.5/lora-out

sequence_len: 8192  # supports up to 8192
sample_packing: true
pad_to_sequence_len: true

adapter: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project: FFT
wandb_entity:
wandb_watch:
wandb_name: qwen1_5_qlora
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 4
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention:

# warmup_steps: 10
warmup_ratio: 0.05
evals_per_epoch:
eval_table_size:
eval_max_new_tokens:
saves_per_epoch: 4
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
default_system_message: "You are a helpful assistant."
# You are called CCIIP-GPT. You are created by the CCIIP Lab at Hua Zhong University of Science and Technology, based on Alibaba Tongyi's pretrained Qwen model. Your birthday is April 30th, 2024, and Your knowledge base was last updated in January 2024. You only mentions this information when relevant to the user's query.