ChatGLM2-6B 微调模型

参考：ChatGLM2-6B-PT

参数

PRE_SEQ_LEN=16
LR=2e-3
NUM_GPUS=1

torchrun --standalone --nnodes=1 --nproc-per-node=$NUM_GPUS main.py \
    --do_train \
    --train_file train_data/train_100k.json \
    --validation_file train_data/dev_1k.json \
    --preprocessing_num_workers 10 \
    --prompt_column question \
    --response_column answer \
    --overwrite_cache \
    --model_name_or_path THUDM/chatglm2-6b \
    --output_dir output/adgen-chatglm2-6b-pt-$PRE_SEQ_LEN-$LR-20230718 \
    --overwrite_output_dir \
    --max_source_length 256 \
    --max_target_length 512 \
    --per_device_train_batch_size 2 \
    --per_device_eval_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --predict_with_generate \
    --max_steps 1000 \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate $LR \
    --pre_seq_len $PRE_SEQ_LEN \
    --quantization_bit 4

train metrics

  epoch                    =        0.2
  train_loss               =     0.3848
  train_runtime            = 3:22:04.09
  train_samples            =      78577
  train_samples_per_second =       1.32
  train_steps_per_second   =      0.082

license: unlicense