Tutorial2- Finetuning on exported feedback data - Exception: Input does not have the expected format

[laifuchicago]

To Author:
I tried your Tutorial2 and get some problems--Exception: Input does not have the expected format. I used your Feedback function, and currently have no idea. I guess the original squad format has "Title". Just like this:

"title" : "xxx",
"paragraphs" : [
{
"context" : "xxx",
"qas" : [
{
"question" : "xxxx",
"id" : "xxxx",
"answers" : [
{
"answer_start" : xxx,
"text" : "xxxx"
}
]
},
But your export feedback does not have. That might be the reason? Do you have any suggestions? Thank you.

Jonathan Sung

[tholor]

Hi @laifuchicago ,

Thanks for reporting this. The exported format did in fact not comply with SQuAD and caused trouble when training with FARM.
Fixed it in #241 .

Please also note:
The context currently contains all text of the indexed document. If you indexed short documents / passages, it will just work fine. However, if you index very long texts, you will have very long texts in the SQuAD "context", which will result in many negative training samples (passages with no annotated answer) and can impact training performance. We will address this in a future PR.

[laifuchicago]

To Author:
Thank you, it can work now!
Jonathan Sung