Update README.md
pipeline is text gen
Actually, this is completely multimodal.
If you pass the token
it will actually respond with either token or
Occasionally it will erroneously produce the token but this can be filtered through innovative use of regex we hope, and future releases should not have this problem.
We have found that in addition to text, this can be used to infer audio or image input.
The model was actually trained extensively on the MSPaint blank dataset as noted in the description, and thus gets approximately 100% of the expected tokens in the test dataset. 1-1 ratio match.
We have a concern of over-fitting for this very reason. But we feel, strongly, that the bubbles in the watercooler are pretty cool, so the loss function probably represents ground truth.