liujch1998 / gkp Goto Github PK

The prompts file such as csqa2_prompts.txt you provided seems only contain 5 context as prompt.
Does all test data use the same prompts?

I reproduce the program in CSQA2.0, but the result shows the model predicts "yes" all the time. So the score of CSQA2.0 in my experiment is just about 50% which is is far lower than the performance in your paper.
And I ran the program according to the instructions:

CUDA_VISIBLE_DEVICES=0 python inference/infer_t5.py
--task csqa2
--model-type allenai/unifiedqa-t5-3b
--input-path data/csqa2/knowledge/context_gpt3.dev.csqa2.json

Can you find the reason of it? Thank you very much.

Hello, I have a problem with line 58 in inference/infer_t5.py

if args.task == 'csqa2' and args.model_ckpt is None:
    scores[1] += 7 

I don't quite understand the purpose of this command line, and the choice of the number 7. Why this operation is only employed in the csqa2 dataset?

Many thanks for your nice codes!

Hello, I'm currently conducting experiments with Generated Knowledge Prompting to reproduce your results.

I have noticed that experimental results on CSQA dataset with the vanilla T5-11B model do not match the reported results on the paper.

The current codebase returns 33.4% accuracy on CSQA dev set, while the paper reports 39.89%.

Are there any missing components or differences between the codebase and the paper?

Thanks.

Hello!
Which UNICORN do you choose? unicorns/lr-2e-3_batch-size-32?