CP-Tuning- model predicted same labels as ground truth label during evaluate and predict mode

[Soumyajain29]

I was trying to run CP tuning code. Code was working fine in training mode. But at inference time (during evaluate and test mode), getting almost 100% accuracy on multiple datasets. The pred.csv file that model saves after evaluation also contains same predicted labels as ground truth labels. However this is not the case when model is evaluated on validation set during training time. Attached the screenshot of the same. I'm unable to figure out where things are getting wrong in the code. Please look into this issue.

This is the script I'm using-

easynlp \
    --app_name=text_classify \
    --mode=evaluate \
    --worker_count=${WORKER_COUNT} \
    --worker_gpu=${WORKER_GPU} \
    --tables=./fewshot_dev.tsv \
    --outputs=pred.tsv \
    --output_schema=predictions \
    --input_schema=sid:str:1,sent1:str:1,sent2:str:1,label:str:1 \
    --worker_count=1 \
    --worker_gpu=1 \
    --first_sequence=sent1 \
    --second_sequence=sent2 \
    --label_name=label \
    --append_cols=sid,label \
    --label_enumerate_values=0,1 \
    --checkpoint_dir=./fewshot_model/ \
    --micro_batch_size=8 \
    --sequence_length=512 \
    --user_defined_parameters="
        enable_fewshot=True
        type=cpt_fewshot
        pattern=sent1,label,with,sent2,summarize
    "

[chywang]

pls look into this issue @ztl-35

[Soumyajain29]

any update on this? @ztl-35 @chywang