[Project]
name=myLLM
version=1.0
user=test
proxies={"http": "127.0.0.1:7890", "https": "127.0.0.1:7890"}

# The pipeline represents a predefined workflow composed of various components in the following context.
pipeline=Trainer,Export,Chat

[TokenizerLoader]
# You can configure all the parameters that appear in the methods `transformers.AutoConfig.from_pretrained`
and `transformers.AutoTokenizer.from_pretrained`.
pretrained_model_name_or_path=THUDM/chatglm2-6b
use_fast=True
;split_special_tokens=True
;padding_side=right
device_map=auto
trust_remote_code=True

[ModelLoader]
# You can specify additional parameters for the class below, for example, `transformers.AutoModelForCausalLM`
class=transformers.AutoModel
pretrained_model_name_or_path=THUDM/chatglm2-6b
print_model_structure=False
trust_remote_code=True
cache_dir=./cache
device_map=auto
use_auth_token=False
;torch_dtype=bf16
use_gradient_checkpointing=False

# for LLama and Falcon models
;rope_scaling=dynamic
model_max_length=2000
flash_attn=False
shift_attn=False
;quantization_bit=8
;double_quantization=4
quantization_type=
finetune_args=FinetuneArguments

[FinetuneArguments]
# type can be full,freeze or lora
type=lora
checkpoint_dir=

# you can set parameters available in the peft.LoraConfig method, make sure to prefix the parameter names with lora_config, for example, `lora_config_target_modules` instead of `target_modules`.
lora_config_task_type=CAUSAL_LM
lora_config_inference_mode=False
lora_config_r=4
lora_config_lora_alpha=32
lora_config_lora_dropout=0.05
lora_config_target_modules=query_key_value
# bias can be set none or all or lora_only
;lora_config_bias=none
;lora_config_fan_in_fan_out=True

;num_layer_trainable=2
;upcast_layernorm=True
;neft_alpha=1e-6

[TrainingArguments]
# You can specify additional parameters for the class below, for example, `transformers.Seq2SeqTrainingArguments`
; class=transformers.TrainingArguments
class=transformers.Seq2SeqTrainingArguments
generation_max_length=256
generation_num_beams=1
per_device_train_batch_size=1
per_device_eval_batch_size=1
gradient_accumulation_steps=2
warmup_steps=20
max_steps=800
;num_train_epochs=1
learning_rate=1e-5
;fp16=True
logging_steps=25
do_train=True
remove_unused_columns=False
output_dir=./output_train
save_safetensors=False
seed=2023

[GenerateArguments]
# you can set parameters available in the transformers.GenerationConfig method
do_sample=True
temperature=1.0
;top_p=1.0
;top_k=50
num_beams=1
max_length=2048
;max_new_tokens=128
repetition_penalty=1.0
length_penalty=1.0

[Trainer]
class=transformers.Seq2SeqTrainer
# The previously defined ModelLoader is referenced here.
model=ModelLoader
# The previously defined TokenizerLoader is referenced here.
tokenizer=TokenizerLoader
# The previously defined TrainingArguments is referenced here.
args=TrainingArguments
# The previously defined DatasetLoader is referenced here.
dataset=DatasetLoader_1
resume_from_checkpoint=False
plot_loss=True
ignore_pad_token_for_loss=True
streaming=False
split_train_val=True
split_train_val_val_size=20
split_train_val_seed=2024
split_train_val_buffer_size=10
steps=train,eval,predict
predict_with_generate=True
ppo_args=PPOArguments
generate_args=GenerateArguments

# stage can be pt,sft,ppo,dpo or rm
stage=sft
# when stage is ppo, you need to set reward_model
reward_model=./output_reward_model

[DatasetLoader_1]
# type should be hf_hub or script or file
path=json
data_files=./data/oaast_sft_zh.json
text_column=instruction
prompt_column=instruction
query_column=input
history_column=history
response_column=output
system_column=system
split=train
cache_dir=./cache
streaming=False
use_auth_token=False

# etl process params
# The previously defined TokenizerLoader is referenced here.
tokenizer=TokenizerLoader
# stage can be pt or sft or rm or ppo
cutoff_len=128
sft_packing=True

# Rendering types for prompts,different value is required for different LLM
render=chatglm2
label_mask_prompt=True

[Export]
# The TokenizerLoader defined above is referenced.
tokenizer=TokenizerLoader
# The ModelLoader defined above is referenced.
model=ModelLoader
output_dir=./export/chatglm2_lora
max_shard_size=5gb

[Chat]
tokenizer=TokenizerLoader
model=ModelLoader
pretrained_model_name_or_path=./export/chatglm2_lora
generating_args=GenerateArguments
render=chatglm2

[ComponentName]
property1=value1
property2=value2
...