Fewshot examples

[TaeyoonKim-DS]

This is an exciting model. How can I add few examples instead of one shot? I am trying to identify brand name from ecommerce product titles such as Amazon.

For example: One shot : Among US Hat Cap Snapback Gift US Seller BRAND, Among US Two shot: [New Adhesion Series 5] Braun Mens Electric Shaver 50-R1200s BRAND, Braun

Any way I can do this?

thanks

[eren23]

I don't know if I understand what is asked here but wouldn't it be possible to use a json file of multiple examples to fit, like this one maybe https://github.com/promptslab/Promptify/blob/main/examples/data/binary.json

Then I think the model would be able to extract brand information from examples like your Two shot above or any other one with arbitrarily longer size and with the general domain knowledge it has should be able to extract any unseen brand as well.

[monk1337]

@ewankim1023, as suggested by @eren23 Please check the examples. you can pass as many examples (shots) as you want using the example parameter.

[monk1337]

@ewankim1023 Here is an example code

You need to format the examples (shots) in [ (example, output) ] format, such as

examples = [("Among US Hat Cap Snapback Gift US Seller", {'T': 'Brand', 'E': 'Among US'}), 
                     ("[New Adhesion Series 5] Braun Mens Electric Shaver 50-R1200s", {'T': 'Brand', 'E': 'Braun'})]

# now giving a new sentence after a few short learning
new_input    = "OnePlus phones are good for multitasking"

result       = nlp_prompter.fit('ner.jinja',
                          domain      = 'ecommerce',
                          text_input  = new_input, 
                          labels      = ['Brand'], 
                          examples = examples)

print(eval(result['text']))

Output:

[{'T': 'Brand', 'E': 'OnePlus'}]