search

nlpodyssey / cybertron

Cybertron: the home planet of the Transformers in Go
BSD 2-Clause "Simplified" License
289 stars 26 forks source link

Training my own model #25

Open ior308 opened 1 year ago

ior308 commented 1 year ago

Hi,

is there a very simple example using spago to train my own models for Q&A ? It is only enough for a single sentence just to understand the process.

Thank you

ior308 commented 1 year ago

Hi, any news ?

ior308 commented 1 year ago

Using Q&A with my own document I got line/cols and scores, and a txt response. Can I get a spoken response adding a generative model trained on the same data ? Can you show me an example ?

Regards

mooijtech commented 1 year ago

Do you mean "text to speech"? That would be out of scope for this project.

ior308 commented 1 year ago

No :) Imagine having a text document describing the Italian Republic Day. Today with the Q&A, you get an extraction of the text that best answers a question and its position. It would instead be possible to generate a generative response based on the question and answer response,, for instance:

"When is Republic Day celebrated?" 9.778867743 { "Answers": [ { "Text": "It is celebrated every year on June 2nd", "Beginning": 143, "End": 177, "Score": 0.688159051418598 }, { "Text": "every year on June 2nd", "Beginning": 156, "End": 177, "Score": 0.2833200460374781 } ] }

Example of generative response: “Republic Day is a day of celebration national... and is celebrated every year on June 2nd.."

This is a response richer in information, although some text is present in the document.