ChatGLM-6B 微调模型

参考：ChatGLM2-6B-PT

使用 kaggle 训练：kaggle notebook 地址

具体参数如下：

!CUDA_VISIBLE_DEVICES=0,1 python3 main.py \
    --do_train \
    --train_file /kaggle/input/sql-create-context-json/train_10k.json \
    --validation_file /kaggle/input/sql-create-context-json/dev_1k.json \
    --prompt_column question \
    --response_column answer \
    --overwrite_cache \
    --model_name_or_path THUDM/chatglm-6b-int4 \
    --output_dir output/adgen-chatglm-6b-int4-pt-128-2e-2 \
    --overwrite_output_dir \
    --max_source_length 64 \
    --max_target_length 64 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 16 \
    --predict_with_generate \
    --max_steps 1000 \
    --logging_steps 10 \
    --save_steps 1000 \
    --learning_rate 2e-2 \
    --pre_seq_len 128 \
    --quantization_bit 4

***** train metrics *****
  epoch                    =        3.2
  train_loss               =     0.1746
  train_runtime            = 7:41:22.34
  train_samples            =      10000
  train_samples_per_second =      1.156
  train_steps_per_second   =      0.036

评估：https://www.kaggle.com/code/javonyan/chatglm-6b-pt-evaluate

predict metrics 
  predict_bleu-4             =    95.6857
  predict_rouge-1            =    96.2341
  predict_rouge-2            =     92.906
  predict_rouge-l            =    95.9385
  predict_runtime            = 0:49:37.03
  predict_samples            =       1000
  predict_samples_per_second =      0.336
  predict_steps_per_second   =      0.168

license: unlicense

language:

tags: